Scaling deep learning for material imaging with a pseudo 3D model for domain transfer Authors: Kunning Tang, Ryan T. Armstrong, Peyman Mostaghimi, Yufu Niu, Quentin Meyer, Chuan Zhao, Donal P. Finegan, Melissa Popeil, Kamaljit Singh, Hannah Menke, Alexandros Patsoukis Dimou, Tom Bultreys, Arjen Mascini, Mark Knackstedt, Ying Da Wang Published: 2025 ABSTRACT The recent introduction of deep learning methods for image processing has greatly advanced the characterization of materials using three-dimensional (3D) X-ray imaging techniques. However, deep learning models often have difficulty performing consistently across images owing to unavoidable variations in imaging conditions, which create inconsistencies even for the same material. As a result, networks must frequently be retrained for new datasets, limiting their applicability and generalization. Thus, it is critical to reduce the variations between images to enable a single model to process multiple datasets. Herein, we introduce P3T-Net, a pseudo-3D domain transfer network that transfers diverse 3D images into a uniform domain before processing using deep learning models. Remarkably, P3T-Net enables the reuse of previously trained networks for processing new images and considerably reduces the computational cost of transferring 3D images across domains. These unique capabilities were demonstrated in the following scenarios: (i) image enhancement of fast scans for geological rock and hydrogen fuel cells, (ii) enhancement of images to match the quality of multi-source imaging for lithium-ion batteries, (iii) accurate segmentation of images captured under different conditions, and (iv) tera-scale 3D transfer (1011 voxels) on a single GPU. Overall, the proposed approach addresses cross-domain inconsistencies across various materials and conditions, thereby enabling more robust and generalizable deep learning solutions for a wide range of material imaging tasks. Introduction 3D imaging techniques such as micro/nano X-ray computed tomography (micro/nano-CT) have revealed the properties of complex and continually evolving materials, such as lithium-ion batteries, hydrogen fuel cells, electrolyzers, and materials that are also challenging to image owing to their heterogeneity, such as geological rocks and critical minerals1–6. The main challenges in characterizing these materials from 3D imaging techniques arise from two factors: (i) limitations in imaging quality, such as the trade-off between scan times and noise-to-signal ratio, and the trade-off between resolution and field of view (FOV)7, and (ii) difficulties in image processing, such as the complexity of multi-phase segmentation6. These challenges have been addressed by the rapid development of deep learning with the proliferation of data-driven and image-based tasks7,8. These tasks include image denoising, which shortens the acquisition time without sacrificing image quality9,10; unbiased multi-phase segmentation, which exceeds manual segmentation in accuracy and speed4,11; and super-resolution, which increases the pixels per feature size, thereby improving detail representation while maintaining an adequate FOV of the materials6,12,13. Although these advancements have led to significant changes in the field, several key challenges remain. These include (i) achieving generalizability across different domains and under diverse conditions, (ii) the restricted learning capacity of 2D networks when applied to 3D images4,14–16, and (iii) addressing the memory limitations and reducing processing time associated with large 3D images, especially those with more than a billion (>109) voxels6,17–19. These challenges arise from the following fundamental issues: (i) domain mismatch between trained networks and inference inputs, which limits their generalizability20–22. This mismatch can arise from both intrinsic material characteristics, such as density, phase fraction, and heterogeneity, and extrinsic imaging factors, including the type of scanning device, scanning conditions, and physical state of the sample during imaging. (ii) The inability of 2D networks to process 3D images owing to information loss along the third axis23,24. Adjacent XY slices in 3D micro/nano-CT images are similar and linearly related in Z-space; however, 2D networks will not inherently understand this Z-space similarity and will output a nonlinear misalignment between these adjacent slices. This results in inaccurate quantification of material properties, such as pore connectivity and material structure. (iii) Incompatibility between memory-intensive 3D networks and handling of large-scale 3D data. Although the utilization of 3D networks can resolve third-axis information, the range of learnable features available in 3D training patches for the networks is restricted owing to the limitations of GPU memory capacity17,25. To address these challenges, advanced deep learning solutions are needed for material characterization that can efficiently handle images from diverse samples scanned under different devices and conditions, as well as be capable of processing large-scale 3D data. The key to solving these tasks lies in adapting them to multiple domains, particularly in managing variations in information distribution, such as grayscale values, image semantics, and styles26,27. Image-to-image translation (I2I) and domain adaptation (DA) are essential techniques for aligning image distributions across multiple domains to address domain mismatches. I2I focuses on transforming images to match the styles or characteristics of different domains28,29, whereas DA adapts trained models to generalize effectively on new domains, ensuring consistent performance despite domain shifts30. Generative adversarial networks (GANs) are commonly used in I2I, including supervised approaches such as Pix2pix31; unsupervised methods such as CycleGAN32, CyCADA33, SCONE-GAN34, and NICE-GAN35; semi-supervised techniques such as Mixmatch36; and models that incorporate GAN with vision transformer architecture, such as ResViT37. For DA, approaches include reducing the discrepancy between domains38,39, employing contrastive learning40,41, diffusion-based methods42,43, and adversarial learning, which share similar objectives with I2I44,45. Both techniques have proven to be effective in aligning images under different distributions. Furthermore, considering that paired data where each image in the source domain is precisely aligned to the target domain is often difficult to obtain, especially for real materials, networks like CycleGAN and CyCADA are typically preferred for domain transfer. These networks can handle images without a one-to-one relationship or precise geometric alignment, known as unpaired data, and offer greater flexibility with fewer constraints on the training data. When transferring unpaired 3D large-scale data across domains, both 2D and 3D networks have their respective strengths and limitations. From a computational perspective, 3D networks are highly memory-intensive and contain more trainable parameters, which makes them challenging to scale when processing large 3D material images46,47. The computational load increases with the size of the data and scales with the third power, which demands significant GPU memory. 2D networks are more scalable and memory efficient. However, 2D networks extract information only from 2D planes, missing spatial relationships along the third axis. This limitation can result in less accurate domain transfer when dealing with volumetric data. Although 3D networks can directly learn 3D structures from the data, preserving the continuity of spatial relationships along the third axis, training such networks to handle 3D nano-/micro-CT images, which typically range from 1000 to 20003 voxels in material imaging, is challenging owing to GPU memory constraints. Consequently, there is a trade-off between computational resources and model performance. Specifically, training a well-performing 3D domain transfer network often requires access to high-performance computing resources, such as H200 GPUs. In contrast, when using more commonly available GPUs that offer 11 to 48 GB of memory, with 24 GB being the most common12,48,49, 3D networks often face long training times and reduced performance owing to the need for smaller training patches or simplified architectures. In unpaired training, it is also important to ensure that the source and target patches share similar characteristics, such as material types and local physical features, to enable effective and robust learning. For 2D networks, this requirement is relatively easy to satisfy because patches can be sampled at or above the representative elementary volume (REV), ensuring adequate feature representation. However, in 3D, this becomes more challenging owing to the limited 3D patch size and the spatial variability of features across different slices within a volumetric patch. Because human experts cannot directly visualize the full 3D structure during sampling, ensuring characteristic similarity between the source and target domains becomes significantly more difficult. We introduce the Pseudo-3D Domain Transfer Network (P3T-Net), an unpaired 3D domain transfer framework that uses two 2D networks to process the X-Y and Z planes, enabling effective 3D domain transfer at tera-large scales (>1010 voxels). By decoupling the 3D transfer task into two 2D perspectives, P3T-Net significantly reduces memory usage and the number of trainable parameters, thereby improving scalability and efficiency. This design supports the use of larger 2D patches, enhancing the network’s ability to learn global representative features, which is particularly important in unsupervised settings with unpaired data. We validated the network on 29 datasets that span diverse real-world materials and imaging conditions in geoscience, energy storage, and fuel cell applications. In addition to achieving strong pixel-wise and image-based accuracy, P3T-Net preserves key physical properties across domains. Finally, we compared its computational cost, training time, and scalability with those of conventional 3D domain transfer networks, demonstrating that P3T-Net can process images a thousand times larger using a single GPU. Results Overview of the pseudo-3D domain transfer network P3T-Net is built on the principal concept of transforming a 3D space using three sequential 2D networks. This approach effectively addresses the third-axis misalignment of 2D networks on 3D data and enhances the scalability of 2D networks for 3D domains, thereby enabling the inference of large 3D domains on a single GPU. An overview of P3T-Net is shown in Fig. 1. P3T-Net consists of three main modules: a semantic indication module, a semantic-consistent domain transfer module, and a misalignment fixing module (Fig. 1a–c). During the training phase, each module was trained independently and sequentially. The process begins with training the semantic indication module, which utilizes 2D grayscale images from the target domain along with their corresponding segmentation ground truths. Next, the semantic-consistent domain transfer module was trained using 2D slices in the X-Y plane from both the source and target domains. In this stage, the previously trained semantic indication module was employed to extract semantic features and ensure consistency across domains. Subsequently, each 2D slice in the X-Y plane from the 3D source domain was processed through the trained domain transfer module and stacked to form the 3D transferred source domain. This transferred volume then serves as the input to the misalignment fixing module, which is trained to correct nonlinear misalignments along the third axis. During the inference phase, each 2D slice in the X-Y plane of the 3D source domain is passed through the semantic-consistent domain transfer module to produce the transferred output. These outputs are stacked to form a 3D volume, from which slices along the Y-Z plane are then passed to the misalignment fixing module to correct misalignment in that direction. Finally, the corrected slices were processed using the semantic indication module and assembled into the final segmented 3D domain. A detailed summary of the inputs and outputs for both the training and inference phases is provided in Tables S1 and S2.Fig. 1Overview of the P3T-Net architecture.It consists of three main modules: (a) Semantic indication module. The latent feature (z) is extracted and used for semantic consistency for the semantic-consistent domain transfer module; b 2D unpaired semantic-consistent domain transfer module based on the CycleGAN architecture. XT and XS are the input images from the target and source domains, respectively. All loss functions are presented in blue font. Semantic consistency is achieved by minimizing the differences between the latent features (z) of the input and output images and the cross-entropy loss between the segmentation of the reconstructed target domain (XT→S→T) and the ground truth. c Misalignment fixing module for fixing third-axis misalignment. d Inference process. Overview of the P3T-Net architecture. It consists of three main modules: (a) Semantic indication module. The latent feature (z) is extracted and used for semantic consistency for the semantic-consistent domain transfer module; b 2D unpaired semantic-consistent domain transfer module based on the CycleGAN architecture. XT and XS are the input images from the target and source domains, respectively. All loss functions are presented in blue font. Semantic consistency is achieved by minimizing the differences between the latent features (z) of the input and output images and the cross-entropy loss between the segmentation of the reconstructed target domain (XT→S→T) and the ground truth. c Misalignment fixing module for fixing third-axis misalignment. d Inference process. With the combination of these modules, the main benefits of P3T-Net are as follows: i) it is designed in a fully unpaired format using a CycleGAN architecture as the bottleneck (Fig. 1b), eliminating the need for paired image registration between training images from the source and target domains. ii) It enhances semantic correctness by minimizing the discrepancy between latent features of the transferred and target domains. These features, which are extracted using a trained semantic indication module (Fig. 1a), guide the optimization process. iii) Finally, all modules are built in 2D but can perform efficient domain transfer in 3D. This was achieved using an additional misalignment fixing module (Fig. 1c). By integrating this module, P3T-Net unlocks 2D scalability on 3D data, allowing it to learn third-axis information. The inference phase of P3T-Net is displayed in Fig. 1d. The detailed training schedule and loss functions can be found in the Methods section. To demonstrate the performance and effectiveness of P3T-Net, we applied it to four materials with various imaging conditions in the following sections. Enhancing laboratory-based micro-CT Laboratory-based micro-CT scanners capture internal material structures at the micrometer scale, but suffer from slow imaging times. Lower photon flux laboratory-based scanners, compared to fast scan sources (e.g., synchrotron), require long scans (1–10 h) to produce images with sufficient quality for further processing and physical quantification. For digital material analysis, it is crucial to scan numerous samples to ensure the accuracy of a study and to sufficiently represent the structure. With limited CT equipment, an effective way to increase the scan capacity is to reduce the scanning time while maintaining acceptable image quality. To examine the effectiveness of P3T-Net in reducing the scanning time for laboratory-based scanners, sandstone rock suitable for geological storage of CO2 and H2 is used50. The sandstone was scanned for approximately 10 h and 7 min. Figure 2a–c shows the 10 h scan-time image as the target domain and its 3D segmentation result, and Fig. 2d, e shows the 7 min scan-time image as the source domain. In this example, these two images exist as voxel-wise pairs, serving as validation data of P3T-Net assessed on a one-to-one domain basis. The image acquisition workflow is described in Section “Materials and Datasets”. Although these two images are paired, training is conducted in an unpaired manner by cropping patches from different regions of the images.Fig. 2Scanning time reduction of the laboratory-based micro-CT image of sandstone using P3T-Net.a X-Y plane, X-Z plane, and Y-Z plane of a long-scan high-quality Bentheimer sandstone image used as the target domain as well as the GT. b 3D visualization of the target domain. c Segmentation results for the target domain. d X-Y plane, X-Z plane, and Y-Z plane of a short-scan low-quality Bentheimer sandstone image used as the source domain. e 3D visualization of the source domain. f Segmentation results of the source domain using the trained semantic indication module. g Example of a transferred domain with misalignment at each plane. Misalignment happens in the X-Z plane and the Y-Z plane. h 3D view of the transferred domain with misalignment. i Segmentation result of the transferred domain using the trained semantic indication module. j Example of the transferred domain after fixing the misalignment on each plane. k 3D view of the transferred domain after misalignment fixing. l Segmentation result of the transferred domain after misalignment fixing using the trained semantic indication module. Scanning time reduction of the laboratory-based micro-CT image of sandstone using P3T-Net. a X-Y plane, X-Z plane, and Y-Z plane of a long-scan high-quality Bentheimer sandstone image used as the target domain as well as the GT. b 3D visualization of the target domain. c Segmentation results for the target domain. d X-Y plane, X-Z plane, and Y-Z plane of a short-scan low-quality Bentheimer sandstone image used as the source domain. e 3D visualization of the source domain. f Segmentation results of the source domain using the trained semantic indication module. g Example of a transferred domain with misalignment at each plane. Misalignment happens in the X-Z plane and the Y-Z plane. h 3D view of the transferred domain with misalignment. i Segmentation result of the transferred domain using the trained semantic indication module. j Example of the transferred domain after fixing the misalignment on each plane. k 3D view of the transferred domain after misalignment fixing. l Segmentation result of the transferred domain after misalignment fixing using the trained semantic indication module. The network’s ability to denoise the 7-minute scan-time image and achieve the same image quality as the 10 h scan-time image is demonstrated in Fig. 2. Figure 2g, h shows that with the semantic-consistent domain transfer module, the 2D X-Y plane slice of the 7 min scan can be transferred into 10-hour quality, while the X-Z and Y-Z planes have notable misalignment. Applying the misalignment fixing module to the Y-Z plane corrects the misalignment. Once the X-Y and Y-Z plane images have been transferred and corrected, the X-Z plane will also be adjusted accordingly (Fig. 2j, k). The transferred domain achieves a close visual match after misalignment fixing compared to the 10 h scan-time image across three planes, which is quantifiable in this case in a paired manner for validation. The Structural Similarity Index Measure (SSIM) and Peak Signal-to-Noise Ratio (PSNR) metrics of the transferred domain compared to the 10 h scan-time image are significantly higher than those of the 7 min scan-time image (Fig. 3a). Moreover, the SSIM achieved by our method even surpasses that of the paired network TomoGAN, which was developed for image denoising10 and details of TomoGAN can be found in Section “Materials and Datasets”. The PSNR values for P3T-Net before and after misalignment fixing were similar; however, the SSIM showed a noticeable improvement after fixing, indicating that the transferred domain with the misalignment fixing module better preserved the third-axis structure and layout of the target domain.Fig. 3Quantitative and physical validation of the domain transferred result for the laboratory-based micro-CT image.a SSIM and PSNR comparison of the unpaired P3T-Net before/after misalignment fixed with the paired method. b Volume fraction comparison of each phase between GT, source domain, P3T-Net with misalignment, and P3T-Net after misalignment fixing. c Surface area comparison of each phase between GT, source domain, P3T-Net with misalignment, and P3T-Net after misalignment fixing. d Pore size distribution comparison of each phase between GT, the source domain, and the transferred image. e Relative permeability simulation on the GT domain, source domain, P3T-Net with misalignment, and P3T-Net with misalignment fixing. kr_n refers to the relative permeability for the non-wetting phase, kr_w refers to the relative permeability for the wetting phase. f Wetting and non-wetting phase flow simulation results of the GT and P3T-Net transferred domain. Quantitative and physical validation of the domain transferred result for the laboratory-based micro-CT image. a SSIM and PSNR comparison of the unpaired P3T-Net before/after misalignment fixed with the paired method. b Volume fraction comparison of each phase between GT, source domain, P3T-Net with misalignment, and P3T-Net after misalignment fixing. c Surface area comparison of each phase between GT, source domain, P3T-Net with misalignment, and P3T-Net after misalignment fixing. d Pore size distribution comparison of each phase between GT, the source domain, and the transferred image. e Relative permeability simulation on the GT domain, source domain, P3T-Net with misalignment, and P3T-Net with misalignment fixing. kr_n refers to the relative permeability for the non-wetting phase, kr_w refers to the relative permeability for the wetting phase. f Wetting and non-wetting phase flow simulation results of the GT and P3T-Net transferred domain. Meanwhile, essential physical parameters for geo-rock research, such as phase connectivity (Euler number), phase surface area, pore size distribution, and two-phase relative permeability curve, have been validated51. The primary goal of conducting 3D imaging is to characterize these parameters. For a comparison of physical accuracy, we directly segmented the source domain using a traditional segmentation workflow (Fig. 2f)52–54, which included non-local means filtering (NLM), 3D watershed segmentation, and morphological opening and closing in the TM Avizo software. P3T-Net results, both before and after misalignment correction, were segmented using a trained semantic indication module (Fig. 2i, l). A visual comparison between the traditional segmentation result and the P3T-Net result is also shown in Fig. S1a. The percentage difference of the Euler number and surface area compared to the GT domain is reported in Fig. 3b, c. This shows that the semantic-consistent domain transfer module significantly improves the physical accuracy, reducing the difference in Euler numbers from the GT domain to the source domain by approximately 4, 25, and 215 times for pores, grains, and high-density phases, respectively. Similar to the surface area, the difference after transfer is reduced by 2-5 times compared to the source domain. When the misalignment fixing module is added, the difference from the GT domain is further minimized for both the Euler number and surface area, achieving an additional 2 to 18 times reduction compared to the result with misalignment. This indicates that the third-axis misalignment caused by the 2D network has a significant influence on the physical analysis and must be fixed. Furthermore, the pore-scale transport characteristics are also analyzed, showing close matches between the GT domain and P3T-Net after misalignment fixing in terms of pore size distributions and relative permeability curves simulated using the Lattice Boltzmann Method55 (Fig. 3d–f). We also analyzed the stability of P3T-Net during training. As shown in Fig. S2a, b, by comparing the PSNR values of a 2D transferred image to the paired slice from the target domain, we observe that the training of the semantic-consistent domain transfer module stabilizes after 80 epochs. For the misalignment fixing module, the PSNR value of a 3D subblock plateaus after eight epochs. Figure S2c, d present visualizations of the 2D transferred image generated by the semantic-consistent domain transfer module and the Y-Z plane of the 3D subblock generated by the misalignment fixing module across various epochs. From this analysis, we can conclude that unlike typical GAN-based networks, where training is unstable and achieving equilibrium between the generator and discriminator is challenging, P3T-Net demonstrates stable training. This stability is attributed to its cycle-based architecture and the use of multiple pixel-to-pixel loss functions, such as cycle consistency loss and semantic consistency loss in the semantic-consistency domain transfer module, as well as pixel consistency loss in the misalignment fixing module, which effectively constrain the training process. Furthermore, since the semantic indication module and semantic consistency loss are closely linked to target domain segmentation, a segmentation-noise sensitivity analysis is conducted to evaluate P3T-Net’s performance under various types of noise in the target domain segmentation. Details of this analysis are provided in Section 1 and Fig. S3–S5 of the Supplementary Materials. We also added a detailed comparative study between P3T-Net and state-of-the-art 3D domain transfer networks, which can be found in Section 2 and Fig. S6 in the Supplementary Materials. Scan time reduction and segmentation of 4D dynamic Images The capabilities of P3T-Net have been extended to process low-quality, high-noise, 4D fast dynamic scans, such as images acquired at synchrotron sources. The transfer from low-quality synchrotron images to a long-scan high-quality domain is challenging because of the trade-off between scanning time (the capability of capturing the dynamic process) and image quality. This case focuses on two-phase flow in porous rocks, which is crucial for underground water management, oil/gas recovery, and underground CO2/hydrogen storage56,57. A sandstone sample, imaged at the Diamond synchrotron (Diamond Light Source, Harwell Science and Innovation Campus, UK) at beamline I13-2, underwent drainage with decane-brine replacing water in the rock. Synchrotron-based images, acquired in 24 s (Fig. 4d–f), are noisy with minimal visual distinction between phases, making traditional segmentation challenging. Accompanying these highly noisy images is an unregistered long-scan image of another wet sandstone sample scanned for 7 h (Fig. 4a–c). The synchrotron image is trained through P3T-Net in conjunction with this unpaired sandstone image. The quality enhancement after misalignment fixing is shown in Fig. 4k–m. This arises from the implementation of all stages of P3T-Net, as seen by comparing Fig. 4h–j with Fig. 4k–m across the X-Z and Y-Z planes. Additional examples of misalignment fixing and final domain transfer results at various drainage stages are shown in Fig. S7 and S8. Aside from recovering the bulk decane and water phase, finer localized features can also be recovered by P3T-Net. An example is given in Fig. S9, where the smaller pores filled with decane are obscured by high-frequency noise in the synchrotron image but become clearly visible after domain transfer. The domain match is demonstrated by the grayscale histograms of the target domain and transferred domain (Fig. 4g) as well as the naturalness image quality evaluator (NIQE) and perception-based image quality evaluator (PIQE) plot of pixel-perceptual quality in Fig. 4n, where the PIQE and NIQE of the source domain shift towards the target domain after transferring.Fig. 4Scan time reduction of the dynamic synchrotron image of sandstone using P3T-Net.a–c X-Y plane, X-Z plane, and Y-Z plane of a long-scan high-quality Bentheimer sandstone image under wet conditions used as the target domain. d–f X-Y plane, X-Z plane, and Y-Z plane of a dynamic synchrotron micro-CT scan of the Bentheimer sandstone used as the source domain. g Grayscale distribution of the target domain, source domain, and transferred domain. h–j X-Y plane, X-Z plane, and Y-Z plane of the transferred synchrotron scan with misalignment. n NIQE-PIQE perceptual quality plot showing how the image quality of transferred images is shifted from the source domain to the target domain. k–m X-Y plane, X-Z plane, and Y-Z plane of the transferred synchrotron scan after misalignment fixing. o Comparison of traditional synchrotron image processing with our P3T-Net processing. p A 2D visualization of the synchrotron image. q Segmentation result using the traditional image processing method. r Segmentation result using the semantic indication module on the transferred image. s Euler number (connectivity) of the decane and water phase for the manual processing result and P3T-Net processing result at the water saturation of 0.59. Scan time reduction of the dynamic synchrotron image of sandstone using P3T-Net. a–c X-Y plane, X-Z plane, and Y-Z plane of a long-scan high-quality Bentheimer sandstone image under wet conditions used as the target domain. d–f X-Y plane, X-Z plane, and Y-Z plane of a dynamic synchrotron micro-CT scan of the Bentheimer sandstone used as the source domain. g Grayscale distribution of the target domain, source domain, and transferred domain. h–j X-Y plane, X-Z plane, and Y-Z plane of the transferred synchrotron scan with misalignment. n NIQE-PIQE perceptual quality plot showing how the image quality of transferred images is shifted from the source domain to the target domain. k–m X-Y plane, X-Z plane, and Y-Z plane of the transferred synchrotron scan after misalignment fixing. o Comparison of traditional synchrotron image processing with our P3T-Net processing. p A 2D visualization of the synchrotron image. q Segmentation result using the traditional image processing method. r Segmentation result using the semantic indication module on the transferred image. s Euler number (connectivity) of the decane and water phase for the manual processing result and P3T-Net processing result at the water saturation of 0.59. Meanwhile, noisy synchrotron images are segmented using a traditional method involving binary watershed segmentation of the dry image, registration to the wet image, binary segmentation of water phases in the wet image, and morphological erosion/dilation to suit user-defined segmentation quality. This process takes approximately 30 h for 20 images (Fig. 4o). For a single experiment, over 200 images are commonly collected, requiring approximately two weeks of manual processing. The traditional segmentation results show under-segmentation of the water phase (Fig. 4q). In contrast to this inaccurate, multi-stage, manual segmentation workflow, P3T-Net achieves a more visually accurate result (Fig. 4r) in just 1.5 h for processing 20 images. A detailed visual comparison of the segmentation results obtained using the traditional method and P3T-Net is shown in Fig. S1b. Unlike the first case, where the sandstone was scanned dry, we captured the dynamic distribution of the two fluids within the sample. Thus, an interesting physical parameter here is the fluid connectivity at different timesteps. The physical accuracy of phase connectivity, measured by the Euler number of the transferred segmented images, shows a negative value for both water and decane at a water saturation of 0.59 (Fig. 4s), indicating that both fluids are well-connected, which is physically reasonable at this water saturation58. This contrasts with the manual result, where decane is significantly disconnected owing to the segmentation error. We also plot the water saturation and Euler number of the decane for all 20 time-series transferred images in Fig. S10a, b. As anticipated, decane is sparsely connected at the beginning but becomes more connected as more water is removed58, whereas water is well-connected at the beginning but becomes disconnected after being replaced by decane. A visualization of the drainage process on a series of transferred images is shown in Fig. S10c. The efficacy of P3T-Net is evident when considering the quality of the fast-time scan. Hardware limitations of X-ray systems have become prominent in capturing the dynamics of fast processes, with traditional image processing unable to produce analyzable images. In contrast, the efficient pseudo-3D architecture produces physically accurate segmentation results within a fraction of the time. A dynamic core flooding scan, which traditionally requires several weeks to process, can now be accomplished in a matter of hours. The results obtained at the synchrotron could thus be processed during the beam time. This represents a substantial shift in the planning and execution of experiments, given the limited time of synchrotron imaging59. Segmentation of out-of-domain cathode nano-CT images Successful domain transfer on images with diverse scan conditions is critical for scans performed with varying speeds and different instruments, requiring collective segmentation and analysis with minimal user bias. This is the key limitation of deep-learned image segmentation models, which are only valid on images acquired with the same imaging conditions, thus requiring both the training of a new model on a case-by-case basis and a human operator to systematically label images. P3T-Net, however, can individually learn the domain mapping of unpaired 3D images. Newly scanned images can be transferred to an existing domain, allowing a segmentation model trained on the existing domain to consistently deliver high-quality segmentation to new images. Avoiding human bias is critical for battery characterization, as variations in user-selected segmentation parameters can induce large uncertainties. P3T-Net mitigates this efficiently by bringing out-of-domain images into the same domain for consistent segmentation. Figure 5 provides a visual comparison of four nano-CT images of lithium-ion battery cathodes, including two mono-model scans (absorption scan), one phase scan, and a dual-mode scan. Dual-mode scanning requires 60% more scan time than mono-mode scanning and a significant expense to install the second mode60. Transferring from mono-model to dual-mode scan quality is crucial for fast and efficient battery characterization. Herein, the dual-mode scan is the target domain (Fig. 5a), with the other three images used as the source domain (first column in Fig. 5c–e).Fig. 5Out-of-domain image transfer and segmentation on the cathode using P3T-Net.a 3D view of the target domain. b Binary segmentation of the target domain using the semantic indication module. (c) from left to right: 3D views of the cathode nano-CT image of SD1, segmentation of SD1 only, transferred domain 1, and Segmentation of the SD1+P3T-Net. (d) from left to right: 3D views of the cathode nano-CT image of SD2, Segmentation of SD2 only, transferred domain 2, Segmentation of the SD2+P3T-Net. (e) from left to right: 3D views of the cathode nano-CT image of SD3, Segmentation on SD3, transferred domain 3, Segmentation of the SD3+P3T-Net. (f) from left to right: magnified regions of SD1, segmentation of SD1 only, and segmentation of SD3+P3T-Net. Out-of-domain image transfer and segmentation on the cathode using P3T-Net. a 3D view of the target domain. b Binary segmentation of the target domain using the semantic indication module. (c) from left to right: 3D views of the cathode nano-CT image of SD1, segmentation of SD1 only, transferred domain 1, and Segmentation of the SD1+P3T-Net. (d) from left to right: 3D views of the cathode nano-CT image of SD2, Segmentation of SD2 only, transferred domain 2, Segmentation of the SD2+P3T-Net. (e) from left to right: 3D views of the cathode nano-CT image of SD3, Segmentation on SD3, transferred domain 3, Segmentation of the SD3+P3T-Net. (f) from left to right: magnified regions of SD1, segmentation of SD1 only, and segmentation of SD3+P3T-Net. An initial attempt to segment the three source domains using the traditional watershed method involved selecting optimal threshold values for each domain, and the results are shown in Fig. S1c. In source domain 1, the electrode phase was underestimated and mistaken for pore space, whereas in source domain 2, the electrode particles were over-segmented and connected. Segmentation in source domain 3 appears visually correct owing to the clearer phase distinction. Overall, traditional watershed segmentation depends heavily on image quality and is subject to user bias and trade-offs regarding phase accuracy. Additionally, to highlight the importance of domain transfer, we performed three segmentation workflows, as listed below:SD only: source images are segmented directly using the segmentation model trained on the target domain.SD+augmentation: the source domain is segmented using a segmentation network trained with target domain images along with data augmentation. Data augmentation includes histogram matching of the target domains using the source domains as references, followed by the random addition of Poisson noise to the images. The results of histogram matching are shown in Fig. S11, and further details on the data augmentation process are provided in Section 4.4.SD+P3T-Net: the source domain is first transferred using P3T-Net and then segmented using the segmentation network trained with the target domain. SD only: source images are segmented directly using the segmentation model trained on the target domain. SD only: source images are segmented directly using the segmentation model trained on the target domain. SD+augmentation: the source domain is segmented using a segmentation network trained with target domain images along with data augmentation. Data augmentation includes histogram matching of the target domains using the source domains as references, followed by the random addition of Poisson noise to the images. The results of histogram matching are shown in Fig. S11, and further details on the data augmentation process are provided in Section 4.4. SD+augmentation: the source domain is segmented using a segmentation network trained with target domain images along with data augmentation. Data augmentation includes histogram matching of the target domains using the source domains as references, followed by the random addition of Poisson noise to the images. The results of histogram matching are shown in Fig. S11, and further details on the data augmentation process are provided in Section 4.4. SD+P3T-Net: the source domain is first transferred using P3T-Net and then segmented using the segmentation network trained with the target domain. SD+P3T-Net: the source domain is first transferred using P3T-Net and then segmented using the segmentation network trained with the target domain. Figure S12 shows the segmentation results and physical validation for SD-only, SD+augmentation, and SD+P3T-Net. The segmentation of SD only showed the poorest performance, followed by SD+augmentation. The most accurate segmentation, both visually and physically, was achieved by SD+P3T-Net. The second column of Fig. 5c–e presents the source domain segmentation results obtained by directly applying the segmentation model trained on the target domain. By transferring these source domains to the target domain condition (third column in Fig. 5c–e) with misalignment fixing (a magnified example of third-axis misalignment is shown in Fig. S13), improved segmentation and image characteristics of the cathode were achieved using the same model (fourth column in Fig. 5c–e). Figure 5f provides a magnified region where cracks are successfully segmented after the transfer (red box), while over-segmentation of the electrode before the transfer is shown in the blue box. Meanwhile, considering that the histogram-matching technique can be used as a domain transfer method, we also apply histogram matching to three source domains using the target domain as a reference and compare it with the P3T-Net result in Fig. 6a–f. The image characteristics, as measured by PIQE and NIQE61,62 in Fig. 6a–c, show distinct differences between the source and target domains, with the transferred domain aligning well with the target domain’s characteristics. However, despite using the histogram-matching technique, the grayscale histograms of the histogram-matched source domains were well aligned with the target domain (Fig. 6d–f), but their PIQE and NIQE distributions remained largely unchanged. This indicates that the histogram matching method can only transfer the grayscale distribution, whereas P3T-Net can effectively transfer the grayscale distribution, image styles, and finer details to the target domain.Fig. 6Quantitative analysis of out-of-domain cathode image transfer with P3T-Net.a–c NIQE-PIQE perceptual quality plot showing how the image quality of transferred images using histogram matching and P3T-Net are shifted from the source domain to target domain for domain 1, 2, 3. d–f Grayscale distribution of target domain, source domain, and transferred domain using histogram matching and P3T-Net of domain 1, 2, 3. g Effective diffusivity values obtained from the target, source, and transferred domains. SD refers to the source domain. Quantitative analysis of out-of-domain cathode image transfer with P3T-Net. a–c NIQE-PIQE perceptual quality plot showing how the image quality of transferred images using histogram matching and P3T-Net are shifted from the source domain to target domain for domain 1, 2, 3. d–f Grayscale distribution of target domain, source domain, and transferred domain using histogram matching and P3T-Net of domain 1, 2, 3. g Effective diffusivity values obtained from the target, source, and transferred domains. SD refers to the source domain. Diffusivity, a crucial parameter for batteries, affects transport efficiency, power capability, and lifespan. We modeled the effective diffusivity for physical validation. On the segmented images of the target and transferred domains, we computed the effective diffusivity of the pore structure to ensure that the transferred domains aligned with the target domain. This consistency is expected for the same homogeneous cathode material. The effective diffusivity match (Fig. 6g) shows that the transferred domain diffusion coefficients (dotted line bars) align with the target domain, whereas the source domains deviate by up to 300%. Scanning time reduction and multi-label segmentation of out-of-domain fuel cell micro-CT images These concepts were extended by applying multi-label segmentation to an out-of-domain fuel cell with a more complex structure. Both a high-quality image of a hydrogen fuel cell, captured in 3 h, and a low-quality image, captured in 10 min, are used. While hydrogen fuel cells are usually imaged slowly owing to the high signal-to-noise ratio of these porous carbon-based materials, this considerably challenges operando measurements, where fast measurements are needed, and pseudo-operando measurements are typically used63. These unpaired hydrogen fuel cell images show that P3T-Net can precondition and perform consistent semantic segmentation on multi-labeled data from different imaging conditions. Figure S14 provides views of the X-Y, X-Z, and Y-Z planes from the 3-hour scan of the fuel cell (target domain) and the 10 min scan (source domain). The effectiveness of the domain transfer and multi-label segmentation built into the architecture of P3T-Net is shown in Fig. 7. The gas diffusion layer of the 10 min scan image is depicted in Fig. 7a–c, illustrating a 2D cross-section, the transferred result using P3T-Net, and the segmentation outcome generated by the semantic indication module. Figure 7d provides a magnified region of the gas diffusion layer where the complex features of the binder, fiber, and pore can be transferred and segmented. Similarly, Figure 7e–g shows the microporous layer (MPL) and the catalyst layer. Figure 7h shows a close-up view of the gas diffusion layer and MPL following domain transfer.Fig. 7P3T-Net performance on the fuel cell dataset.a–c 2D cross-section of the gas diffusion layer from the source domain, P3T-Net transferred result, and the segmentation result using the semantic indication module. d Magnified region from the red box, highlighting the domain transfer of the gas diffusion layer. e–g 2D cross-section of the gas diffusion layer and MPL from the source domain, P3T-Net transferred result, and the segmentation result using the semantic indication module. h Magnified region from the blue box, highlighting the domain transfer of the gas diffusion layer and MPL. i X-Z plane of the source domain and its transferred result. j Y-Z plane of the source domain and its transferred result. k The NIQE-PIQE scores plot shows how the image quality of the transferred images shifts from the source domain to the target domain. l Grayscale distribution of target domain, source domain, and transferred domain. P3T-Net performance on the fuel cell dataset. a–c 2D cross-section of the gas diffusion layer from the source domain, P3T-Net transferred result, and the segmentation result using the semantic indication module. d Magnified region from the red box, highlighting the domain transfer of the gas diffusion layer. e–g 2D cross-section of the gas diffusion layer and MPL from the source domain, P3T-Net transferred result, and the segmentation result using the semantic indication module. h Magnified region from the blue box, highlighting the domain transfer of the gas diffusion layer and MPL. i X-Z plane of the source domain and its transferred result. j Y-Z plane of the source domain and its transferred result. k The NIQE-PIQE scores plot shows how the image quality of the transferred images shifts from the source domain to the target domain. l Grayscale distribution of target domain, source domain, and transferred domain. The images of the source, P3T-Net output, and segmentation visually demonstrate the achievable segmentation improvement, as the source-domain image quality is insufficient for multi-label segmentation. We also manually segmented the source domain image using traditional methods, including grayscale thresholding and the Magic Wand tool in Avizo (FEI)4. The manual and P3T-Net results are shown in Fig. S1d. Visual inspection shows significant errors in the manual segmentation of the fiber, binder, and MPL phases, whereas P3T-Net accurately segments these phases despite a similar grayscale distribution and high noise levels. Additionally, Fig. 7i, j show that the X-Z and Y-Z planes of the source are transferred to the target domain without third-axis misalignment. Furthermore, Fig. 7k shows a perceptual quality plot of NIQE-PIQE, emphasizing the transition in image quality from the source to the target domain. Finally, Fig. 7l shows the grayscale distributions for the target, source, and transferred domains, highlighting the differences between the source and target domains, and the similarity between the transferred and target domains. Large scale 3D domain transfer approaching tera-scale P3T-Net addresses the computational intensity and low scalability of 3D deep learning methods64,65, which limits the deployability of most deep learning methods for large 3D datasets. To perform domain transfer on these images with 3D consistency instead of 2D consistency, which otherwise results in third-axis misalignment, P3T-Net applies two sequential 2D CNNs: first on the X-Y plane, where most domain difference occurs, and then a CNN on the Y-Z plane to resolve misalignment caused by nonlinear transformations in the X-Y plane. Once the X-Y and Y-Z plane images have been transferred and fixed, the X-Z plane is adjusted accordingly. This approach allows P3T-Net to scale in 2D during inference, enabling GPU resources to infer slice-by-slice rather than requiring the entire 3D domain. To highlight the substantial reduction in computational cost, we compared P3T-Net with the 3D version of CycleGAN, a widely utilized domain transfer network23,66. This comparison focused on five key aspects: training patch size, representative features, model size, GPU usage during inference, and inference time. Both networks were trained using the same target and source domains, as shown in Section “Scan Time Reduction and Segmentation of 4D Dynamic Images”. The computational comparison results are shown in Fig. 8.Fig. 8Computational efficiency and scalability comparison between the P3T-Net and 3D CycleGAN during training and inference stages.a Patch size and corresponding GPU memory usage for training 3D CycleGAN and P3T-Net, with a batch size of 1 for all configurations. For 3D CycleGAN, the patch size is given in cubic dimensions, whereas for P3T-Net, it is in square dimensions. AMP stands for Automatic Mixed Precision, which automatically handles the management of mixed precision between float32 and float16. b Captured FOV with varying patch sizes for the source and target domains using examples from Section “Scan Time Reduction and Segmentation of 4D Dynamic Images”. c Domain size and corresponding GPU memory usage for 3D CycleGAN and P3T-Net during inference. d Visualization of the domain transfer results of 3D CycleGAN and P3T-Net. e 3D visualization of tera-scale large-domain transfer result using P3T-Net. f Inferring time of P3T-Net, including domain transfer, misalignment fixing, and total time (domain transfer + misalignment fixing). Computational efficiency and scalability comparison between the P3T-Net and 3D CycleGAN during training and inference stages. a Patch size and corresponding GPU memory usage for training 3D CycleGAN and P3T-Net, with a batch size of 1 for all configurations. For 3D CycleGAN, the patch size is given in cubic dimensions, whereas for P3T-Net, it is in square dimensions. AMP stands for Automatic Mixed Precision, which automatically handles the management of mixed precision between float32 and float16. b Captured FOV with varying patch sizes for the source and target domains using examples from Section “Scan Time Reduction and Segmentation of 4D Dynamic Images”. c Domain size and corresponding GPU memory usage for 3D CycleGAN and P3T-Net during inference. d Visualization of the domain transfer results of 3D CycleGAN and P3T-Net. e 3D visualization of tera-scale large-domain transfer result using P3T-Net. f Inferring time of P3T-Net, including domain transfer, misalignment fixing, and total time (domain transfer + misalignment fixing). We trained the 3D CycleGAN on a 24 GB consumer-grade GPU. Our choice of consumer-grade GPU for benchmarking is intended to reflect a more commonly available scenario in academic research, providing a realistic baseline for the computational constraints faced by many researchers. As shown in Fig. 8a, a patch size of less than 1103 voxels with a batch size of one reaches the memory limit for 3D CycleGAN. Applying the quantization technique (Automatic Mixed Precision, AMP) to the 3D CycleGAN increases the maximum patch size to 1363 voxels. In contrast, for P3T-Net, which is entirely 2D-based, the patch size can easily reach 5122 pixels, using only 9.7 GB of GPU memory without AMP and 5.4 GB with AMP. The patch size is crucial for training, as illustrated in Fig. 8b. For this specific sample, a patch smaller than 1282 pixels contains limited representative features of the sample, and these features can be substantially different between patches from the source and target domains. For instance, at a size of 962 pixels, the patch from the source domain contains more pores, whereas the patch from the target domain primarily consists of fluid. Training 3D networks with these mismatched patches can lead to the generation of artifacts or hallucinations in the transferred images. Additionally, the model size and number of trainable parameters for 3D CycleGAN and P3T-Net were compared, as detailed in Tables S4 and S5. 3D CycleGAN has a model size that is 21 times larger than P3T-Net with a patch size of 962 pixels, and the number of trainable parameters is 1.6 times greater. The model sizes of 3D CycleGAN and P3T-Net were compared using different training patch sizes, as illustrated in Fig. S15. The model size of 3D CycleGAN increases sharply from 0.6 to 29 GB as the input patch size increases from 323 to 1283 voxels, whereas P3T-Net’s size increases steadily from 0.2 to 0.9 GB. This dramatic increase in the model size for 3D CycleGAN requires significantly more GPU memory and demonstrates its low scalability, which limits its applicability to large-scale 3D image processing. A computational efficiency comparison during inference between P3T-Net and 3D CycleGAN showed that P3T-Net can transfer a domain with a size of over 4000 × 4 000 × 5000 voxels in a single inference. Fig. 8c shows the memory scalability of the network compared to 3D CycleGAN, with P3T-Net only reaching the GPU memory limit of 24 GB at a domain size of 40003+ voxels, whereas CycleGAN fails to reach 2563 voxels in a single inference. Figure 8d shows a visualization of the single inference result of a 2003 subdomain from 3D CycleGAN and P3T-Net. Although 3D CycleGAN transfers the source image to the target image, artifacts are evident, as highlighted in the yellow box, where pore spaces are converted into solid phases after the transfer. This observation aligns with our discussion of Fig. 8b, which attributes the issue to the limited patch size and representative features used for training the 3D model. This limitation leads to feature inconsistencies across patches, ultimately causing the generation of artifacts or hallucinations in the transferred images. The NIQE is calculated as a perceptual quantification of the transferred image. A higher NIQE value (6.9) was observed for 3D CycleGAN, whereas P3T-Net achieved a score of 5.1, which closely aligned with the target image score of 4.9. This indicates that P3T-Net transfers source images more effectively to the target domain in terms of perceptual quality. Figure 8e shows a terascale transferred domain using P3T-Net in a one-time inference manner. The inference time scalability of P3T-Net (Fig. 8f is also shown for domain sizes up to the memory limit, with parabolic scaling indicative of the 2D scalability of the method. The X-Y domain transfer and misalignment fixing stages scale additively, requiring 3.8 h to transfer a large-scale 40003 voxel domain using an Nvidia RTX 4090. It is also worth mentioning that such a 3D domain in Fig. 8e would require approximately 10,400 inferences on non-overlapped subcubes (20 × 20 × 26 subblocks of 2003 voxels) for a 3D CycleGAN-type neural network, taking approximately 15 h, excluding accounting for extra time to apply methods to overcome boundary-overlap artifacts, such as using overlapping blocks, applying local padding, or training extra networks to fix boundary discontinuities. Discussion This study introduces P3T-Net, a pseudo-3D domain transfer network that leverages efficient 2D architectures to accurately transfer large-scale 3D imaging data, overcoming the memory and scalability limitations of existing 3D approaches. P3T-Net was rigorously validated on 29 real-world 3D datasets across diverse materials and imaging conditions, demonstrating strong domain transfer and practical applicability. Its effectiveness is further highlighted through four engineering case studies from geoscience, energy, and materials domains, addressing critical challenges such as enhancing 4D imaging quality, reducing scan times, consistent segmentation, and improving characterization in battery and fuel cell systems. We also demonstrate the computational efficiency and tera-scale scalability of P3T-Net on a workstation-grade computer, a key factor determining the deployability of the network in processing large 3D images. The grayscale distribution shifts and image quality differences arising from variations in material density, phase fraction, imaging conditions, and scan timing were addressed using a CycleGAN-based architecture that transfers imaging styles while preserving semantic consistency through its semantic indication module. To handle phase heterogeneity, P3T-Net uses a pseudo-3D architecture that allows larger in-plane training patches, enabling better representation of both global features and local fluctuations across slices, which is critical for accurately transferring features in highly heterogeneous materials, such as fuel cells. Although P3T-Net requires less human involvement during training, some conditions are necessary prior to training: (i) materials from the target and source domains should exhibit similar characteristics, such as material type (e.g., the material type for both domains should be sandstone, rather than sandstone as the target domain and carbonate as the source domain) and sample condition (e.g., dry or wet). A standard assumption of P3T-Net is that the source and target domains should contain shared cross-representative features. Maintaining these similarities is essential to ensure that the domain transfer is meaningful and that the transferred images accurately preserve the structural and compositional properties of the materials. (ii) Target and source domains should have similar pixels per average feature size (resolution) to ensure that they contain comparable FOV, and (iii) unpaired training patches sampled from target and source domains must contain similar features. Despite being designed for unpaired training, maintaining feature similarity is essential for stable network performance, especially for discriminators. For example, in fuel cell images, the model might fail if training patches for the target domain are exclusively from the gas diffusion layer, while those for the source domain are from the membrane and catalyst layer. Consistent sourcing is crucial to avoid training instability. This is also why staying in 2D for unpaired domain transfer is preferred because feature consistency is more difficult to achieve in 3D space owing to more complex 3D features and limited training patch sizes. With similar features presented in target and source training slices, P3T-Net supports domain transfer in arbitrary directions, meaning that any directional anisotropy (e.g., diagonal structures) can be captured and transferred. In terms of future work, although P3T-Net has been demonstrated for single-source-to-single-target domain mapping, it is also capable of transferring multiple source domains to the same target domain (Fig. S16). However, it cannot handle multiple target domains. Addressing this requires incorporating target-class conditioning into the pipeline by linking the target-domain class to the source input to appropriately guide the model. Another direction is to explore the adaptability of P3T-Net to other imaging modalities, such as MRI, medical CT, and industrial X-rays, which will require understanding modality-specific imaging physics, grayscale semantics, and domain mismatches. The third direction is to assess the robustness of the model under large domain mismatches or unseen material types. While P3T-Net has demonstrated the ability to handle significant mismatches, it may struggle when there is no shared or cross-representative feature space between the source and target domains. In such ill-posed scenarios, the task shifts from domain transfer to generative synthesis, where no deterministic mapping can be guaranteed. Future research could explore strategies to better characterize or quantify the degree of domain mismatch prior to transfer, potentially incorporating feature variation estimation to identify cases in which meaningful transfer is not possible. Lastly, although P3T-Net effectively computes bulk physical properties and captures localized features, it remains limited in recovering rare features that appear infrequently in the target domain. This limitation can be addressed by incorporating user-defined annotations, such as through the semantic indication module, or by designing additional localized loss functions or attention mechanisms to guide the model toward these features. This would extend P3T-Net’s applicability to rare-event imaging tasks, such as defect detection and failure analysis. Methods Materials and datasets To convey the applications of domain transfer and the associated limitations (third-axis misalignment and 3D inefficiency), we selected four types of cases to demonstrate unpaired denoising, image enhancement, and segmentation consistency on a large scale in 3D. The reason we selected these cases is that 3D imaging techniques have been widely applied to understand these materials on digital images, and many challenges in terms of 3D image processing exist that pose difficulties in advancing digital materials characterization1,6,51. These include (i) a sandstone rock imaged with different speeds (10 min and 3 h) on a bench-top micro-CT, (ii) lithium-ion battery cathodes scanned under different conditions on a bench-top nano-CT scanner, (iii) highly noisy synchrotron scans of sandstone under two-phase flow, and (iv) a proton exchange membrane fuel cell imaged at different qualities on a bench-top micro-CT. Table 1 provides the scanning details and image information. The estimated scanning duration was determined by considering the exposure time, number of projections, accumulation settings, dark- and bright-field scans, and triggering.Table 1Materials used to test the proposed network and the imaging and scanning information of each 3D dataMaterialImaging techniqueScan durationExposure timeProjectionsAccumulationsDomain sizeVoxel sizeSandstone (target)Micro-CT9 h 27 min0.55 s2880201560 × 1560 × 12802.15 μmSandstone (source)Micro-CT7 min0.55 s36011560 × 1560 × 12802.15 μmLi-ion cathode (source1)Nano-CT (absorption)7 h25 s9011344 × 336 × 331128 nmLi-ion cathode (source2)Nano-CT (absorption)4.6 h15 s9011332 × 364 × 342128 nmLi-ion cathode (source3)Nano-CT (phase)9.8 h36 s9011356 × 352 × 316128 nmLi-ion cathode (target)Nano-CT (dual-mode)––––335 × 356 × 340128 nmSandstone (source)Synchrotron micro-CT24 s0.02 s8001760 × 760 × 10243.28 μmSandstone (target)Micro-CT6 h 45 min0.45 s288016520 × 540 × 25606.75 μmFuel cell (target)Micro-CT3 h0.55 s28805800 × 800 × 1162.25 μmFuel cell (source)Micro-CT10 min0.55 s3601800 × 800 × 1162.25 μm Table 1Materials used to test the proposed network and the imaging and scanning information of each 3D dataMaterialImaging techniqueScan durationExposure timeProjectionsAccumulationsDomain sizeVoxel sizeSandstone (target)Micro-CT9 h 27 min0.55 s2880201560 × 1560 × 12802.15 μmSandstone (source)Micro-CT7 min0.55 s36011560 × 1560 × 12802.15 μmLi-ion cathode (source1)Nano-CT (absorption)7 h25 s9011344 × 336 × 331128 nmLi-ion cathode (source2)Nano-CT (absorption)4.6 h15 s9011332 × 364 × 342128 nmLi-ion cathode (source3)Nano-CT (phase)9.8 h36 s9011356 × 352 × 316128 nmLi-ion cathode (target)Nano-CT (dual-mode)––––335 × 356 × 340128 nmSandstone (source)Synchrotron micro-CT24 s0.02 s8001760 × 760 × 10243.28 μmSandstone (target)Micro-CT6 h 45 min0.45 s288016520 × 540 × 25606.75 μmFuel cell (target)Micro-CT3 h0.55 s28805800 × 800 × 1162.25 μmFuel cell (source)Micro-CT10 min0.55 s3601800 × 800 × 1162.25 μm Materials used to test the proposed network and the imaging and scanning information of each 3D data The first case, where we used sandstone rock imaged with different speeds on a bench-top micro-CT, was used as the benchmark case for validating P3T-Net performance. During the imaging of the sandstone rock for the fast scan (source domain) and slow scan (target domain), the sample was consistently placed on the rotation stage, with no human intervention during either scan. Fast and slow scans were achieved by adjusting the number of projections, accumulations, and exposure time, which are parameters related to the X-ray source. Under these conditions, the fast and slow scans were naturally paired, capturing the sample in exactly the same state and coordinates. Thus, although P3T-Net was trained in an unpaired manner by cropping patches from different regions of the fast and slow scans, during inference, the transferred fast scan had a corresponding paired slow-scan ground truth for benchmarking. A schematic workflow for scanning the benchmark case is shown in Fig. S17. In this case, the domain transfer tends to denoise the fast scan to the same image quality as the slow scan. Because paired datasets are available, a paired denoising network, TomoGAN, was used for comparison with P3T-Net. TomoGAN is a modified version of the U-Net architecture67, with fewer downsampling and upsampling layers, zero-padding applied to all convolutional layers, and the use of three stacked adjacent input slices. A key strength of TomoGAN lies in its use of adjacent slices, which enhances the reconstruction quality by leveraging spatial continuity. Because adjacent slices often exhibit strong correlations-sharing similar shapes with slight variations in size-stacking them as multiple input channels in the first CNN layer effectively acts like applying a multivariate linear filter, helping to suppress noise. P3T-Net architecture, loss functions, and training schedules Semantic indication module Semantic indication module (Fig. 1a) embeds a U-ResNet4,51 with residual connections for latent feature extraction and the subsequent image segmentation purposes. The detailed architecture of the U-ResNet can be found in Fig. S18. The network consists of an encoder and a decoder. The encoder aims to extract multilevel latent features from the inputs, whereas the decoder transfers the latent features back to the pixel level for image classification/segmentation. The residual connection incorporated in U-ResNet is based on the concept of the residual network68, which enables mitigation of the degradation and overfitting issues. The objective function of the semantic indication module (Lsi) is a multi-label categorical cross-entropy loss, which can be formulated as1\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$${L}_{{{{\rm{si}}}}}({I}_{GT},\hat{I})=-{\sum }_{i=1}^{n}{I}_{GT,i}\,\log ({\hat{I}}_{i})$$\end{document}Lsi(IGT,I^)=−∑i=1nIGT,ilog(I^i)where n is the number of classes, IGT denotes the ground truth label, \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\hat{I}$$\end{document}I^ represents the predicted label from U-ResNet. The semantic indication module is the first component to be trained in P3T-Net using the target domain and its corresponding segmented ground truth. Once trained, it is employed to extract semantic information during the training process of the semantic-consistent domain-transfer module. Semantic-consistency domain transfer module The semantic-consistent domain transfer module (Fig. 1b) primarily consists of two parts: a domain transfer network built upon the CycleGAN architecture, and a semantic consistency part that uses semantic features extracted from the trained semantic indication module. The domain transfer module is constructed based on the original CycleGAN32, comprising two generators (GS→T and GT→S) and two discriminators (DS and DT). Generators are built on a ResNet architecture68 with five residual blocks, and discriminators consist of four convolutional blocks (each including a convolutional layer, Instance Normalization layer, and LeakyReLU activation function), followed by a convolutional layer that reduces the feature channels to 1, and an average pooling layer with flattening. The generator (GS→T) and discriminator (DS) constitute the cycle of domain transfer from the source domain to the target domain and back to the source domain (from XS to XS→T and to XS→T→S). The generator (GT→S) and discriminator (DT) form the cycle of domain transfer from target to source and back to target (from XT to XT→S and to XT→S→T). To learn the mapping between the source and target domains, we first applied the adversarial loss to the cycle from the source domain to the target domain, expressed as2\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$${L}_{{{{\rm{GAN}}}}}({G}_{{{{\rm{S}}}}\to {{{\rm{T}}}}},{D}_{{{{\rm{T}}}}})= {{\mathbb{E}}}_{{X}_{T} \sim {p}_{{{{\rm{data}}}}}({X}_{T})}\left[\log {D}_{{{{\rm{T}}}}}({X}_{T})\right]\\ +{{\mathbb{E}}}_{{X}_{S} \sim {p}_{{{{\rm{data}}}}}({X}_{S})}\left[\log \left(1-{D}_{{{{\rm{T}}}}}({G}_{{{{\rm{S}}}}\to {{{\rm{T}}}}}({X}_{S}))\right)\right]$$\end{document}LGAN(GS→T,DT)=EXT~pdata(XT)logDT(XT)+EXS~pdata(XS)log1−DT(GS→T(XS))where XS and XT denote samples from the source domain and target domain. In the Eq. (2), the first term represents the expectation over real target domain images XT sampled from the data distribution XT ~ pdata(XT), while the second term represents the expectation over domain transferred images XS→T = GS→T(XS), where XS is the source domain images sampled from the data distribution XS ~ pdata(XS). In practice, the expectations in Eq. (2) are estimated by averaging over mini-batches of target and domain transferred samples. The adversarial training objective is then formulated as the following min-max optimization problem:3\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$${G}_{{{{\rm{S}}}}\to {{{\rm{T}}}}}^{*},\ {D}_{{{{\rm{T}}}}}^{*}=\arg {\min }_{{G}_{{{{\rm{S}}}}\to {{{\rm{T}}}}}}\ {\max }_{{D}_{{{{\rm{T}}}}}}\ {L}_{{{{\rm{GAN}}}}}({G}_{{{{\rm{S}}}}\to {{{\rm{T}}}}},{D}_{{{{\rm{T}}}}}).$$\end{document}GS→T*,DT*=argminGS→TmaxDTLGAN(GS→T,DT).In the meantime, the generator GT→S transforms target images XT into fake source images (XT→S). DS attempts to distinguish between XT→S and XS. A similar adversarial loss for this cycle can be expressed as:4\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$${L}_{{{{\rm{GAN}}}}}({G}_{{{{\rm{T}}}}\to {{{\rm{S}}}}},{D}_{{{{\rm{S}}}}})= {{\mathbb{E}}}_{{X}_{S} \sim {p}_{{{{\rm{data}}}}}({X}_{S})}\left[\log {D}_{{{{\rm{S}}}}}({X}_{S})\right]\\ +{{\mathbb{E}}}_{{X}_{T} \sim {p}_{{{{\rm{data}}}}}({X}_{T})}\left[\log \left(1-{D}_{{{{\rm{S}}}}}({G}_{{{{\rm{T}}}}\to {{{\rm{S}}}}}({X}_{T}))\right)\right]$$\end{document}LGAN(GT→S,DS)=EXS~pdata(XS)logDS(XS)+EXT~pdata(XT)log1−DS(GT→S(XT))In the Eq. (4), the first term represents the expectation over real source domain images XS sampled from the data distribution XS ~ pdata(XS), while the second term represents the expectation over domain transferred images XT→S = GT→S(XT), where XT is the target domain images sampled from the data distribution XT ~ pdata(XT). The adversarial training objective is then formulated as the following min-max optimization problem:5\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$${G}_{{{{\rm{T}}}}\to {{{\rm{S}}}}}^{*},{D}_{{{{\rm{S}}}}}^{*}=\arg {\min }_{{G}_{{{{\rm{T}}}}\to {{{\rm{S}}}}}}\ {\max }_{{D}_{{{{\rm{S}}}}}}\ {L}_{{{{\rm{GAN}}}}}({G}_{{{{\rm{T}}}}\to {{{\rm{S}}}}},{D}_{{{{\rm{S}}}}})$$\end{document}GT→S*,DS*=argminGT→SmaxDSLGAN(GT→S,DS)A one-fold adversarial loss is not sufficient to ensure individual mapping between the source and target domains. Thus, a cycle consistency loss is introduced to strengthen the learned function, ensuring that each unique image in the source domain can be transferred to a corresponding image in the target domain and reconstructed. The cycle consistency loss can be expressed as6\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$${L}_{{{{\rm{cyc}}}}}({G}_{{{{\rm{S}}}}\to {{{\rm{T}}}}},{G}_{{{{\rm{T}}}}\to {{{\rm{S}}}}})= {{\mathbb{E}}}_{{X}_{S} \sim {p}_{{{{\rm{data}}}}({X}_{S})}}\left[\left\Vert {X}_{S\to T\to S}-{X}_{S}\right\Vert \right]\\ +{{\mathbb{E}}}_{{X}_{T} \sim {p}_{data({X}_{T})}}\left[\left\Vert {X}_{T\to S\to T}-{X}_{T}\right\Vert \right]$$\end{document}Lcyc(GS→T,GT→S)=EXS~pdata(XS)XS→T→S−XS+EXT~pdata(XT)XT→S→T−XTWhere XS→T→S and XT→S→T denote the reconstructed images of the source and target domains, where XS→T→S = GS→T(GT→S(XS)) and XT→S→T = GT→S(GS→T(XT)), respectively. In addition, we introduced a semantic consistency loss (Lsc) to preserve semantic consistency using the latent features extracted from the trained semantic indication module. The semantic consistency loss enforces semantic alignment among the source, target, and transferred domains (both XS→T and XT→S). (Lsc) can be expressed as7\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$${L}_{{{{\rm{sc}}}}}= {L}_{{{{\rm{sc}}}}}({X}_{S},{X}_{S\to T})+{L}_{{{{\rm{sc}}}}}({X}_{T},{X}_{T\to S})+{L}_{{{{\rm{si}}}}}({I}_{{{{\rm{GT}}}}},{\hat{I}}_{{X}_{T\to S\to T}})\\= {{\mathbb{E}}}_{{X}_{S} \sim {p}_{{{{\rm{data}}}}}({X}_{S})}\left[{\left\Vert {Z}_{S}-{Z}_{S\to T}\right\Vert }^{2}\right]+{{\mathbb{E}}}_{{X}_{T} \sim {p}_{{{{\rm{data}}}}}({X}_{T})}\left[{\left\Vert {Z}_{T}-{Z}_{T\to S}\right\Vert }^{2}\right] \\+ {L}_{{{{\rm{si}}}}}({I}_{{{{\rm{GT}}}}},{\hat{I}}_{{X}_{T\to S\to T}}).$$\end{document}Lsc=Lsc(XS,XS→T)+Lsc(XT,XT→S)+Lsi(IGT,I^XT→S→T)=EXS~pdata(XS)ZS−ZS→T2+EXT~pdata(XT)ZT−ZT→S2+Lsi(IGT,I^XT→S→T).where ZS, ZT, ZT→S, and ZS→T are the latent features extracted from the semantic indication module from XS, XT, XT→S, and XS→T, respectively. \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$${\hat{I}}_{{X}_{T\to S\to T}}$$\end{document}I^XT→S→T denotes segmentation of the reconstructed target domain as shown in Fig. 1b. Misalignment fixing module The misalignment fixing module (Fig. 1c) consists of a trained generator (GS→T) from the semantic-consistency domain transfer module and a 2D GAN designed to correct spatial misalignment. The GAN consists of a generator (GMF) and discriminator (DMF). GMF is constructed using the ResNet architecture68 with four residual blocks, whereas DMF has the same architecture as DS. The trained GS→T generates a fake target image (XS→T), which is then transposed and passed to the misalignment-fixing generator (GMF). The corresponding adversarial loss is defined as8\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$${L}_{{{{\rm{MF}}}}}({G}_{{{{\rm{MF}}}}},{D}_{{{{\rm{MF}}}}})= \,{{\mathbb{E}}}_{{X}_{T} \sim {p}_{{{{\rm{data}}}}}({X}_{T})}\left[\log {D}_{{{{\rm{MF}}}}}({X}_{T}^{{\prime} })\right]\\ +{{\mathbb{E}}}_{{X}_{S\to T} \sim {p}_{{{{\rm{data}}}}}({X}_{S\to T})}\left[\log \left(1-{D}_{{{{\rm{MF}}}}}({X}_{S\to T}^{{\prime} })\right)\right].$$\end{document}LMF(GMF,DMF)=EXT~pdata(XT)logDMF(XT′)+EXS→T~pdata(XS→T)log1−DMF(XS→T′).where \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$${X}_{S\to T}^{{\prime} }$$\end{document}XS→T′ and \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$${X}_{T}^{{\prime} }$$\end{document}XT′ denote the transposed (orthogonal) planes of the source domain after transfer using the trained GS→T and of the target domain, respectively. Additionally, we incorporated a pixel consistency loss (Lpc) to ensure that, following misalignment fixing of the orthogonal planes, the original plane remains consistent with the input. This loss can be formulated as9\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$${L}_{{{{\rm{pc}}}}}={{\mathbb{E}}}_{{X}_{S\to T} \sim {p}_{{{{\rm{data}}}}({X}_{S\to T})}}\left[\left\Vert {({G}_{{{{\rm{MF}}}}}({X}_{S\to T}^{{\prime} }))}^{{\prime} }-{X}_{S\to T}\right\Vert \right]$$\end{document}Lpc=EXS→T~pdata(XS→T)(GMF(XS→T′))′−XS→THere, \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$${X}_{S\to T}^{{\prime} }$$\end{document}XS→T′ is the XS→T after transpose. \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$${({G}_{{{{\rm{MF}}}}}({X}_{S\to T}^{{\prime} }))}^{{\prime} }$$\end{document}(GMF(XS→T′))′ has the same corresponding orientation as XS→T. Overall, the loss objective of P3T-Net comprises two parts: the semantic consistency domain transfer module and the misalignment fixing module. The loss function for the semantic-consistent domain transfer module is formulated as10\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$${L}_{{{{\rm{dt}}}}}= {L}_{{{{\rm{GAN}}}}}({G}_{{{{\rm{S}}}}\to {{{\rm{T}}}}},{D}_{{{{\rm{T}}}}},{X}_{{{{\rm{S}}}}},{X}_{{{{\rm{T}}}}})+{L}_{{{{\rm{GAN}}}}}({G}_{{{{\rm{T}}}}\to {{{\rm{S}}}}},{D}_{{{{\rm{S}}}}},{X}_{{{{\rm{T}}}}},{X}_{{{{\rm{S}}}}})\\ +{\lambda }_{{{{\rm{cyc}}}}}({L}_{{{{\rm{cyc}}}}}({G}_{{{{\rm{S}}}}\to {{{\rm{T}}}}},{G}_{{{{\rm{T}}}}\to {{{\rm{S}}}}}))+{\lambda }_{{{{\rm{sc}}}}}{L}_{{{{\rm{sc}}}}}$$\end{document}Ldt=LGAN(GS→T,DT,XS,XT)+LGAN(GT→S,DS,XT,XS)+λcyc(Lcyc(GS→T,GT→S))+λscLscwhere λcyc is set to 10 and λsc is also set to 0.8. Additionally, the objective of the misalignment fixing module is formulated as follows:11\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$${L}_{{{{\rm{mf}}}}}={\lambda }_{{{{\rm{MF}}}}}{L}_{{{{\rm{MF}}}}}+{\lambda }_{{{{\rm{pc}}}}}{L}_{{{{\rm{pc}}}}}$$\end{document}Lmf=λMFLMF+λpcLpcwhere λMF and λpc are both set to 1 in this study. Note that we did not use the identity loss proposed in the original CycleGAN because the semantic consistency loss we propose already ensures the preservation of grayscale values before and after domain transfer. Training details and schedules During the training stage, an unpaired 2D source (XS) and target images (XT) are input to CycleGAN (GS−T and GT−S), and the transferred source (XS−T) and target (XT−S) images can be obtained. However, with a GAN-based network, image-to-image transfer is stochastic. This poses challenges in digital material image processing because essential material information may be lost or inaccurately represented in the transferred images. To maintain the consistency of the image information, we designed a semantic indication module and used it to extract semantic information during domain transfer training. The semantic indication module was designed as an n-phase segmentation network with the U-ResNet architecture (Fig. S18). It is trained using 2D slices from the target domain (XT) and their corresponding segmented ground truth in a paired manner. The detailed generation of the segmentation ground truth is discussed in the next section. The semantic indication module requires training each time for any new task to learn image semantics. After training the semantic indication module, XT, XT−S, XS, and XS−T are fed to the trained semantic indication module to obtain 1D latent features ZT, ZT−S, ZS, and ZS−T of size 1 × 10, respectively. Semantic consistency loss was then calculated for these latent features. Additionally, the XT−S−T that was transferred back to the target domain from XT−S is also input for the trained semantic indication module to obtain its segmentation result. This process was optimized by considering the cross-entropy loss between the segmentation result and the ground truth of XT. All semantic consistency losses were incorporated with the standard CycleGAN losses (cycle consistency loss, adversarial loss, and identity loss) to guarantee the transfer of both domain characteristics and material information. After training the semantic-consistent domain transfer module, a misalignment fixing module is designed and trained to fix the third-axis misalignment of the output from the semantic-consistent domain transfer module (as shown in Fig. 1c). During training, an input 3D XS with shape [D, H, W] is treated as a 2D batch of consecutive images, with the 3D XS reshaped to 2D images (\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$${X}_{{S}^{{\prime} }}$$\end{document}XS′) with batch size D such that the shape of \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$${X}_{{S}^{{\prime} }}=[D,1,H,W]\equiv [N,C,H,W]$$\end{document}XS′=[D,1,H,W]≡[N,C,H,W], where N is the number of batch size, and C is the channel size. This allows \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$${X}_{{S}^{{\prime} }}$$\end{document}XS′ to be passed directly into the trained 2D GS→T to obtain XS→T with a third-axis misalignment. XS→T is then transposed from [D, 1, H, W] to [H, 1, D, W] to obtain \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$${X}_{S\to T}^{{\prime} }$$\end{document}XS→T′ which is then passed to the misalignment fixing generator (GMF) for fixing the misalignment. GMF is optimised by calculating the adversarial loss between the GMF output with the original XT and the pixel loss between the XS→T and the transposed GMF output (transpose from [H, 1, D, W] back to [D, 1, H, W]). Detailed training steps of the misalignment fixing module can be found in Fig. S19. During inference (Fig. 1d), each 2D slice in the X-Y plane from the 3D source domain is passed through GS−T in the semantic-consistent domain transfer module. After stacking all the 2D output images into a transferred 3D domain, each slice in the X-Z plane of the transferred 3D domain is input to GMF in the misalignment fixing module, resulting in a finalized 3D transferred domain with misalignment fixing. Additionally, if the final objective is to generate the segmentation result, the 3D transferred domain can be directly passed to the trained semantic indication module to generate the final 3D segmentation output. In terms of the training schedule, the segmentation indication module is trained using the Adam optimizer. The training starts with an initial learning rate of 1e−5 and a batch size of 16. The training spans 50 epochs, during which the learning rate is halved whenever the loss plateaus for five epochs. Subsequently, together with the segmentation indication module pre-trained in the first stage, we trained the semantic-consistent domain transfer module with the unpaired images from the target and source domains, using the Adam optimizer with a batch size of 8. The training starts with an initial learning rate of 3e−5 and the same decay strategy is applied. Overall, this stage was trained for 150 epochs. Finally, we trained the misalignment fixing module using the images from the target domain and the inferred images from the second stage using the Adam optimizer with a batch size of 1. The training starts with an initial learning rate of 3e−5 and the same decay strategy is applied. This stage was trained for 30 epochs. To train the semantic-consistent domain transfer module, 2D patches were sampled, whereas 3D patches were cropped to train the misalignment correction module. Specifically, we collected 1800 2D patches with a size of 2242 pixels and 900 3D patches with a size of 643 voxels for the laboratory-based micro-CT case (Section “Enhancing Laboratory-based micro-CT”), 2500 2D patches with a size of 2242 pixels and 600 3D patches with a size of 803 voxels for the synchrotron case (Section “Scan Time Reduction and Segmentation of 4D Dynamic Image”), 2100 2D patches with a size of 2242 pixels and 600 3D patches with a size of 803 voxels for the lithium-ion cathode case (Section “Segmentation of Out-of-Domain Cathode Nano-CT Images”), and 2160 2D patches with a size of 2242 pixels and 600 3D patches with a size of 843 voxels for the fuel cell case (Section “ Scanning Time Reduction and Multi-Label Segmentation of Out-of-Domain Fuel Cell micro-CT Images”). All these training patches were randomly cropped from the 3D target and source domains. Note that for random cropping, it is generally advisable to use patch sizes that are close to or larger than the REV of the material. In imaging, the REV represents the smallest image size of a material that can be considered representative of its bulk properties. When the training patch size meets or exceeds the REV, each randomly cropped patch is more likely to capture statistically representative microstructural features. Consequently, patches from the source and target domains are more likely to contain similar features, which helps to maintain consistency in the learning process. To evaluate this in the selected patch sizes, we conducted REV analysis for all four cases to determine whether the patch sizes used satisfy the REV requirement. Specifically, the REV analysis was based on different structural properties: porosity of the sandstone rock in Case 1, volume fraction of decane in Case 2, porosity of the lithium-ion cathode in Case 3, and porosity of the gas diffusion layer in the fuel cell in Case 4. For each case, patches of varying sizes were randomly sampled from the segmented 3D target domains, and the corresponding property values were computed. Overall, 1,000 patches of each size were randomly sampled for this analysis. As shown in Fig. S20, the distribution of points indicates that as the 2D patch size increases, the calculated values converge toward the global value obtained from the 3D domain. The point distribution reaches a plateau (estimated REV) at a size of approximately 2002 pixels for cases 1, 2, and 3, and at 2242 pixels for case 4. The patch size selected for training the semantic-consistent domain transfer module was 2242 pixels, which meets or exceeds the estimated REV for all cases. In addition to random cropping, several alternative sampling strategies can be considered. These include: (1) human-guided sampling, where patches from the source and target domains are selected to satisfy specific physical constraints. This ensures that the sampled regions are physically meaningful and representative, thereby improving the consistency of the training data across domains. (2) Similarity-based sampling, in which a pre-trained network is used to extract features from patches, and those with similar feature representations are paired across domains. It enhances latent feature-level alignment between patches from the target and source domains, which can reduce domain discrepancy; and (3) block-based sampling, where both the source and target 3D domains are divided into sub-blocks, and patches are sampled from blocks exhibiting similar characteristics, potentially with human guidance. It also enables structured and localized comparisons across the target and source domains to ensure that the patches contain similar features. P3T-Net was developed and tested in PyTorch on a local workstation equipped with a single NVIDIA RTX 4090 GPU (24 GB memory), AMD 3990 CPU, and 128 GB of RAM. The training times for the semantic-consistent domain transfer module and the misalignment fixing module in each case were as follows: 6.7 h for the laboratory-based micro-CT case, 8.9 h for the synchrotron case, 7.8 h for the lithium-ion cathode case, and 8.3 h for the fuel cell case. Training image preparation for semantic indication module The semantic indication module was trained with the target domain and its corresponding ground-truth segmentation. The preparation of the ground truth for the target domain is straightforward, given that these domains typically comprise high-quality, low-noise 3D images. Semantics are defined by the classes segmented from these images. Specifically, in case 1, the semantics of the sandstone rock target image include pores, grains, and high-density minerals. In case 2, the semantics of the dynamic synchrotron sandstone target image consist of water, decane, and solid. It is important to note that although the material is the same in Cases 1 and 2 (sandstone rock), the semantics can vary significantly based on human-defined classifications and the specific research focus. Additionally, for case 3, the semantics of the lithium-ion battery target image consist of the electrode and pore, and in case 4, the semantics of the fuel cell include the pore, fiber, binder, MPL, and catalyst layer. The workflow for generating the target domain segmentation across the four cases is detailed in Table 2. For cases 1–3, the target domains were first preprocessed using the Non-Local Means (NLM) filter to remove noise. Subsequently, multiphase watershed segmentation is applied to cluster voxels into phases, which are defined as semantics. Watershed segmentation was performed using the Avizo software. Cases 1 and 2 were segmented into three phases and post-processed by opening and closing the segmented voxels to remove segmentation errors, especially at the phase boundaries. This is performed using morphological dilation and erosion functions in MATLAB. Finally, in case 4, the segmentation of the fuel cell is substantially more challenging than in other cases, owing to its complex and heterogeneous microstructure and low material contrast, which makes it difficult to accurately differentiate and segment the various components. We used Trainable Weka Segmentation (TWS)69, a machine learning-based segmentation tool, to segment a few selected 2D slices from the entire 3D domain. In the TWS process, we first annotated a subset of 2D slices to generate training labels, which were then used to train a machine-learning classifier. Once trained, the classifier was applied to segment the entire region across all slices. We then manually corrected segmentation errors on the 2D slices using the connected component removal function in MATLAB.Table 2Image processing and segmentation workflow for processing target domains for all casesTarget domainPreprocessingSegmentation softwareSegmentation methodPostprocessingCase 1 (Section “Enhancing Laboratory-based micro-CT”)Sandstone rock (9 h 27 min scan)NLMTM AvizoWatershed (3D)Opening+ClosingCase 2 (Section “Scan Time Reduction and Segmentation of 4D Dynamic Images”)Wet sandstone rock (6 h 45 min scan)NLMTM AvizoWatershed (3D)Opening+ClosingCase 3 (Section “Segmentation of Out-of-Domain Cathode Nano-CT Images”)Li-ion cathode (dual-mode)NLMTM AvizoWatershed (3D)NoneCase 4 (Section “Scanning Time Reduction and Multi-Label Segmentation of Out-of-Domain Fuel Cell micro-CT Images”)Fuel cell (3 h scan)NoneFiji ImageJTWS (2D)Manual correctionNLM: Non-local means noise filter, TWS: Trainable Weka Segmentation, Opening/Closing: morphological dilation and erosion for noise removal. Table 2Image processing and segmentation workflow for processing target domains for all casesTarget domainPreprocessingSegmentation softwareSegmentation methodPostprocessingCase 1 (Section “Enhancing Laboratory-based micro-CT”)Sandstone rock (9 h 27 min scan)NLMTM AvizoWatershed (3D)Opening+ClosingCase 2 (Section “Scan Time Reduction and Segmentation of 4D Dynamic Images”)Wet sandstone rock (6 h 45 min scan)NLMTM AvizoWatershed (3D)Opening+ClosingCase 3 (Section “Segmentation of Out-of-Domain Cathode Nano-CT Images”)Li-ion cathode (dual-mode)NLMTM AvizoWatershed (3D)NoneCase 4 (Section “Scanning Time Reduction and Multi-Label Segmentation of Out-of-Domain Fuel Cell micro-CT Images”)Fuel cell (3 h scan)NoneFiji ImageJTWS (2D)Manual correctionNLM: Non-local means noise filter, TWS: Trainable Weka Segmentation, Opening/Closing: morphological dilation and erosion for noise removal. Image processing and segmentation workflow for processing target domains for all cases NLM: Non-local means noise filter, TWS: Trainable Weka Segmentation, Opening/Closing: morphological dilation and erosion for noise removal. Data augmentation for segmentation network training For the battery cathode material, the nano-CT images used were relatively similar, with the primary difference being the exposure time during scanning. For further comparison with the P3T-Net results, a data augmentation workflow was proposed for the segmentation network, as shown in Fig. S21. The two main differences among the four nano-CT images are the grayscale distribution and noise level. To incorporate these two differences in data augmentation, a histogram-matching technique is applied to the target domains using each source domain as the reference image. This histogram matching is performed in MATLAB using the ‘imhistmatch’ function with the ‘polynomial’ matching method. The target domain after histogram matching has the same grayscale distribution as the source domain. By doing this, three additional target domains are generated (aligned with the three source domains) as the data augmentation results. In the next step, we trained a U-ResNet segmentation network (the same as the model used in the semantic indication module) using the original target domain image along with three target domain images obtained after histogram matching with three source domains. Before training, we considered that different domains have varying levels of noise, with Poisson noise being the most prevalent in nano-CT imaging70. Therefore, we randomly selected 50% of the training patches and added noise to them. A total of 3200 paired training patches, with 800 cropped from the original target domain and an additional 2400 patches generated from the three target domains after histogram matching, served as data augmentation. The trained U-ResNet was then applied to segment the source domains. Apart from data augmentation, the histogram-matching technique can also be used as a domain transfer method. To test its performance, we applied the histogram-matching technique to all three source domains using the target domain as a reference, as shown in Fig. S22. Subsequently, three histogram-matched source domains were obtained and compared with our P3T-Net transferred results. Pixel-wise and physical evaluating metrics Image quality and perceptual index PSNR, a typical pixel-wise criterion, has become the most prevalent method for assessing image fidelity. The PSNR between the domain-transferred image and its ground truth was calculated to evaluate the model’s performance in domain transfer from a pixel-wise perspective. The PSNR can be expressed as12\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$${{{\rm{PSNR}}}}(g,f)=10\,{\log }_{10}\left(\frac{25{5}^{2}}{\frac{1}{mn}\mathop{\sum }_{i=1}^{m}\mathop{\sum }_{j=1}^{n}{\left(g(i,j)-f(i,j)\right)}^{2}}\right)$$\end{document}PSNR(g,f)=10log1025521mn∑i=1m∑j=1ng(i,j)−f(i,j)2where g is the domain-transferred image, f is the ground truth image, m and n are the dimension sizes of a given 2D image. i and j are the pixel indices. SSIM is another classical method for measuring the similarity between two images71. Unlike PSNR, which focuses on pixel-wise information, SSIM evaluates the perceptual quality of images that includes luminance, contrast, and structure. The SSIM formula can be expressed as13\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$${{{\rm{SSIM}}}}(g,f)=\frac{\left(2{u}_{g}{u}_{f}+{c}_{1}\right)\,\left(2{\sigma }_{gf}+{c}_{2}\right)}{\left({u}_{g}^{2}+{u}_{f}^{2}+{c}_{1}\right)\,\left({\sigma }_{g}^{2}+{\sigma }_{f}^{2}+{c}_{2}\right)}$$\end{document}SSIM(g,f)=2uguf+c12σgf+c2ug2+uf2+c1σg2+σf2+c2where g and f are the domain-transferred image and the ground truth, respectively. ug and uf are the means of the domain-transferred image and ground truth, respectively. σg and σf are the standard deviations of the domain-transferred image and the ground truth, respectively. σgf is the covariance of the domain-transferred image and ground truth. c1 and c2 are the regularization constants for luminance and contrast, respectively: c1 = (0.01×L)2, c2 = (0.03×L)2, where L is the dynamic range of the pixel values. For the 8-bit images in all our cases, L = 255. The SSIM is computed locally for each pixel in an image by comparing it to other pixels within an 11 × 11 × 11 window. Then, all window SSIMs were averaged to obtain a global SSIM value. In addition to PSNR and SSIM, NIQE62 and PIQE72, as two non-reference image quality metrics, are introduced to estimate the image quality score of the source, target, and transferred domains. NIQE is designed to measure the naturalness of an image. It is assumed that natural images follow certain statistical distributions, and any deviation from these distributions indicates a reduction in image quality. PIQE focuses more on the human perception of the image. It aims to quantify the perceptual quality of an image by identifying and evaluating image distortions, artifacts, noise, and local variations. In most instances, PIQE performs comparably to NIQE. PIQE assesses blockwise distortion by measuring the local variance of the visibly distorted blocks to calculate the quality score. Generally, lower NIQE and PIQE scores indicate better image quality. Note that PIQE and NIQE are non-reference metrics pretrained on a database of known distortions and pristine images. However, for micro/nano-CT images, subtle textural details and features might be considered perceptible distortions, leading to higher PIQE and NIQE scores. Therefore, NIQE and PIQE are not suitable direct indicators of micro/nano-CT image quality. Instead, they were used to reflect the differences between domains in terms of grayscale values, noise levels, and image style. Histogram distribution The histogram distribution is an approximation method used to represent the distribution of voxel values by counting the number of data points in a bin. Histograms were applied to compare the voxel value distributions between the source, target, and transferred domains. Volume fraction of phases The volume fraction of each phase was calculated from the segmented image. It is defined as the ratio of the number of voxels clustered in each phase to the total number of voxels in the 3D image. Pore size distribution The pore size distribution was obtained by calculating the equivalent diameter of pores in the segmented image. First, each isolated individual pore is labeled from the 3D image using the Separating Object function in TM Avizo. Because the geometry of each pore is irregular, each pore is approximated as a sphere with the same volume. The equivalent diameter of each isolated pore was then calculated and counted to estimate the pore size distribution. Surface area and phase connectivity The surface area and connectivity of each phase as geometric properties are measured from Minkowski functionals73,74. We calculated the surface area of each phase in the images from the source domain, target domain, and transferred domain to evaluate the model’s performance. The equation of the surface area can be simply expressed as14\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$A({X}^{{\prime} })={\int}_{\delta {X}^{{\prime} }}ds$$\end{document}A(X′)=∫δX′dswhere A is the surface area, \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$${X}^{{\prime} }$$\end{document}X′ is the objective embedded in Euclidean space (ω), \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\delta {X}^{{\prime} }$$\end{document}δX′ is the entire surface, ds is a surface element on \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$${X}^{{\prime} }$$\end{document}X′. The number of directions used for estimating the surface area was set to 13. Another Minkowski functional, the Euler number (χ), was introduced to measure the connectivity of the phase. The Euler number formula can be written as15\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\chi ({X}^{{\prime} })=N-L+O$$\end{document}χ(X′)=N−L+O Where N is the number of isolated objects, L is the number of redundant loops, and O is the number of cavities (voids). The voxel-to-voxel connectivity was set to 6 in this study, indicating that only the six neighboring voxels at a distance of 1 were considered connected. The surface area and χ were measured using the Minkowski functional package in MATLAB73. Effective diffusivity The effective diffusivity of a porous domain can be defined based on Poisson’s equation; specifically, it is based on Fick’s law of diffusion75,76. The key consideration in the case of resolved porous structures is the geometry of the domain, that is, the configuration of the pore space and solid space. An elliptic diffusion equation was solved for the segmented void space of the micro/nano-CT images. The conservation of mass can be expressed as:16\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\nabla \cdot {{{\bf{v}}}}=q$$\end{document}∇⋅v=qwhere v is the diffusion velocity field and q represents the source or sink term, which is assumed to be zero in this case. Each voxel is assigned a local diffusion coefficient D, which relates the local diffusion velocity to the concentration gradient as17\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$${{{\bf{v}}}}=-D\nabla C$$\end{document}v=−D∇Cwhere C denotes concentration. Combining equations (16) and (17) gives the Diffusion Equation:18\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$-\nabla \cdot (D\nabla C)=0$$\end{document}−∇⋅(D∇C)=0 This equation is solved with prescribed constant concentration (Dirichlet) boundary conditions on the inlet and outlet using two-point flux approximation (TPFA) and Finite Volume Methods with an Algebraic Multi-Grid (AMG) solver. All solid voxels were removed from the system of equations, resulting in a smaller system matrix. Once the spatial concentration distribution and diffusion velocity profile are calculated, the effective diffusivity is estimated from19\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$${D}_{{{{\rm{eff}}}}}=\frac{{J}_{{{{\rm{D}}}}}{N}_{x}R}{\Delta C}$$\end{document}Deff=JDNxRΔCThe length of the system is expressed as the number of voxels multiplied by the voxel size of the image R. The number of voxels in the main flow direction is shown as Nz, JD is the overall diffusion flow rate, and ΔC is the concentration difference across the domain imposed as boundary conditions. Two-phase flow relative permeability The Navier-Stokes continuum transport equations are estimated using the Lattice Boltzmann Method (LBM). In this study, the Lattice Boltzmann for Porous Media (LBPM) software package55 is used, which models immiscible two-phase flow with interfacial tension and wettability interactions between the fluid and solid phases. No-slip boundary conditions were applied to the fluid-solid boundary, and periodic boundary conditions were applied at the prescribed inlet and outlet faces. Fluids are driven by a uniform body force applied to both phases to mimic co-injection experiments, and an accelerated steady-state simulation protocol is utilized based on the coupling of morphological operations with LBPM77. The relative permeability of a domain is calculated as the ratio of the phase permeability to the absolute permeability at a given fluid saturation:20\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$${K}_{{r}_{{{{\rm{w}}}},{{{\rm{nw}}}}}}({S}_{w})=\frac{{K}_{{{{\rm{w}}}},{{{\rm{nw}}}}}({S}_{w})}{{K}_{a}}$$\end{document}Krw,nw(Sw)=Kw,nw(Sw)Kawhere Kr is the relative permeability, w and nw are the wetting and non-wetting phases, respectively, Ka is the absolute permeability, and Sw is the wetting fluid saturation. In the LBM simulation, the effective phase permeability was calculated using Darcy’s Law21\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$${K}_{{r}_{{{{\rm{w}}}},{{{\rm{nw}}}}}}({S}_{w})=-\frac{{\mu }_{{{{\rm{w}}}},{{{\rm{nw}}}}}\,{\overline{{{{\bf{v}}}}}}_{{{{\rm{w}}}},{{{\rm{nw}}}}}}{\nabla P}$$\end{document}Krw,nw(Sw)=−μw,nwv¯w,nw∇Pwhere μw,nw is the phase viscosity, \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$${\overline{{{{\bf{v}}}}}}_{{{{\rm{w}}}},{{{\rm{nw}}}}}$$\end{document}v¯w,nw is the mean phase velocity, and ∇ P is the system pressure gradient, proportionate to the body force. Reporting summary Further information on research design is available in the Nature Portfolio Reporting Summary linked to this article. Supplementary information Supplementary Information File Reporting Summary Transparent Peer Review file Supplementary Information File Reporting Summary Transparent Peer Review file