DiffSyn: A Generative Diffusion Approach to Materials Synthesis Planning

Elton Pan, Soonhyoung Kwon, Sulin Liu, Mingrou Xie, Alexander J. Hoffman, Yifei Duan, Thorben Prein, Killian Sheriff, Yuriy Roman-Leshkov, Manuel Moliner, Rafael Gomez-Bombarelli, Elsa Olivetti · 2025

Read DiffSyn: A Generative Diffusion Approach to Materials Synthesis Planning by Elton Pan, Soonhyoung Kwon, Sulin Liu, Mingrou Xie, Alexander J. Hoffman, Yifei Duan, Thorben Prein, Killian Sheriff, Yuriy Roman-Leshkov, Manuel Moliner, Rafael Gomez-Bombarelli, Elsa Olivetti (2025) in the Tower of Babel, a free online library of public-domain works preserved in clean, searchable text. No sign-up. The full text is served as a plain-text file at a predictable URL, where text is present: https://antarctic-labs.com/tower-of-babel/arxiv-2509-17094.txt.

Introduction

Elton Pan and a large team plan crystal synthesis with a generative model in a 2025 materials paper. Making zeolites and similar crystals means searching a vast space of recipes. One structure can arise from many recipes, which suits a generative approach. Their model, DiffSyn, trained on tens of thousands of recipes from fifty years of literature.

As a proof of concept, the team made a target material from model-proposed routes. Theory calculations backed the chosen routes. The product had a silicon-to-aluminum ratio higher than any prior record for that material. Higher ratios should improve heat stability. Learned recipe planning could cut years of trial and error in materials labs.

Catalog record

CREATORElton Pan, Soonhyoung Kwon, Sulin Liu, Mingrou Xie, Alexander J. Hoffman, Yifei Duan, Thorben Prein, Killian Sheriff, Yuriy Roman-Leshkov, Manuel Moliner, Rafael Gomez-Bombarelli, Elsa Olivetti
YEAR2025
COLLECTIONPAPERS
FORMATTXT
FILE SIZE73 KB
VERSION1.0
LICENSEhttp://creativecommons.org/licenses/by/4.0/
RIGHTSCC BY
SOURCEarXiv

Source: https://arxiv.org/abs/2509-17094

Browse the full library · About the Tower of Babel

Full text

The opening of the work, from the library's hosted text:

DiffSyn: A Generative Diffusion Approach to Materials Synthesis Planning Elton Pan, Soonhyoung Kwon, Sulin Liu, Mingrou Xie, Alexander J. Hoffman, Yifei Duan, Thorben Prein, Killian Sheriff, Yuriy Roman-Leshkov, Manuel Moliner, Rafael Gomez-Bombarelli, Elsa Olivetti Abstract The synthesis of crystalline materials, such as zeolites, remains a significant challenge due to a high-dimensional synthesis space, intricate structure-synthesis relationships and time-consuming experiments. Considering the one-to-many relationship between structure and synthesis, we propose DiffSyn , a generative diffusion model trained on over 23,000 synthesis recipes spanning 50 years of literature. DiffSyn generates probable synthesis routes conditioned on a desired zeolite structure and an organic template. DiffSyn achieves state-of-the-art performance by capturing the multi-modal nature of structure-synthesis relationships. We apply DiffSyn to differentiate among competing phases and generate optimal synthesis routes. As a proof of concept, we synthesize a UFI material using DiffSyn -generated synthesis routes. These routes, rationalized by density functional theory binding energies, resulted in the successful synthesis of a UFI material with a high Si/Al ICP {}_{\text{ICP}} of 19.0, which is expected to improve thermal stability and is higher than that of any previously recorded. keywords Generative models, Materials synthesis, Porous materials 1 Introduction Materials discovery lays the foundation for modern technologies, from catalysis to electronics butler2018machine . Recent large-scale computational searches of chemical composition and structures merchant2023scaling ; sriram2024open ; zhu2024uncovering ; kim2021solid have uncovered millions of potentially stable, synthesizable materials ( what to synthesize) merchant2023scaling ; zeni2023mattergen ; barroso2024open ; saal2013materials . However, finding viable synthesis routes remains a bottleneck in materials discovery ( how to synthesize) he2023precursor ; huo2022machine ; karpovich2023interpretable ; mahbub2020text because there are many synthesis parameters (composition, conditions, etc.) that interact in complex ways. Moreover, the compute required for atomistic simulations scales poorly with system size, precluding accurate modeling of the underlying physical phenomena in complex materials (i.e., thermodynamics and kinetics) bianchini2020interplay . Consequently, there is continued interest in machine learning (ML) approaches to directly learn from experimental synthesis data to predict materials synthesis parameters at lower computational cost huo2022machine ; karpovich2023interpretable ; pan2024zeosyn . Materials synthesis prediction presents a unique challenge for ML for several reasons. First, structure-synthesis relationships are one-to-many —i.e., a single target structure may form through multiple possible synthesis recipes. Second, the inverse relationship (synthesis-structure) is also one-to-many—i.e., a single recipe may result in the formation of a mixture of products (competing phases) due to the complex interplay of thermodynamic and kinetic pathways bianchini2020interplay . Capturing this phase competition is crucial to selectively synthesize single-phase materials instead of mixtures. Third, complex non-linear interactions exist between synthesis parameters, such as temperature and time (Fig. A10 ) karpovich2023interpretable , requiring approaches that model joint probabilities across multiple synthesis parameters. Predictions must capture relationships among variables to make tradeoffs between parameters leveraging physical information about materials synthesis (e.g., crystallization kinetics). Previous ML approaches to predict synthesis have predominantly used regression approaches karpovich2023interpretable ; prein2024reaction ; schwalbe2023inorganic ; luo2022mof , which deterministically map a representation (e.g., composition huo2022machine ; karpovich2023interpretable , structural features schwalbe2023inorganic , graphs prein2024reaction ) of a material to its synthesis parameters. These approaches are limited because the deterministic mapping is incompatible with the one-to-many nature of structure-synthesis relationships and assume independence between synthesis parameters (Fig. A4 ) pan2024chemically ; karpovich2021inorganic . These factors limit the predictive accuracy of regression approaches and motivate a shift to generative models, which can sample a complex distribution that accounts for the non-linear interactions between parameters in high-dimensional synthesis space. To address these challenges, we introduce a diffusion model for materials synthesis. Diffusion models are a powerful class of generative models that were demonstrated to generate novel, high-quality images conditioned on text saharia2022photorealistic ; ramesh2021zero . Diffusion models can be guided at each step of the denoising process toward a specific objective (e.g., target material) ho2022classifier ; dhariwal2021diffusion . Unlike generative adversarial networks that suffer from mode collapse zhang2018convergence , diffusion models can generate diverse outputs because they are trained to denoise data (e.g., a synthesis route) that have been corrupted with noise. Other generative approaches, namely variational autoencoders and normalizing flows, have limited expressivity due to one-step decoding and affine invertible layers, respectively. In contrast, the iterative denoising process renders diffusion models highly expressive, which enables high sample quality dhariwal2021diffusion . This high expressivity may enable diffusion models to capture boundaries in synthesis space between competing phases. We propose DiffSyn , a diffusion approach to materials synthesis prediction and demonstrate it on zeolites, which are crystalline, microporous materials with applications in catalysis, adsorption, and ion exchange. Zeolite synthesis is challenging due to its high dimensionality (Fig. 1 a), with numerous variables influencing the synthesis outcome (Fig. A1 ). Moreover, multiple modes of valid synthesis routes exist for a given structure (Fig. A5 ). Progress in zeolite synthesis has focused on trial-and-error experiments guided by domain heuristics moliner2019machine . We demonstrate the first application of a guided diffusion model for materials synthesis prediction, which exhibits state-of-the-art performance compared to regression-based and other deep generative approaches. We show that the performance of DiffSyn arises from its ability to capture the one-to-many and multi-modal nature of structure-synthesis relationship in materials. We experimentally validate our approach by synthesizing the UFI zeolite based on DiffSyn -generated synthesis routes. We rationalize these routes using density functional theory (DFT) calculations of inorganic cations that guide UFI synthesis. Together, these results indicate that DiffSyn learns the underlying chemistry that influences synthesis outcomes implicitly from published synthesis recipes. Figure 1: DiffSyn: A diffusion approach to materials synthesis planning . (a) Materials, such as zeolites, often have an one-to-many structure-synthesis relationship, where a structure c_{\text{zeo}} can be synthesized via multiple possible synthesis routes in a high-dimensional synthesis space (compositions \{x_{\text{comp}}\} and conditions \{x_{\text{cond}}\} ). (b) Overview of DiffSyn : Given a desired zeolite structure c_{\text{zeo}} and an organic structure directing agent c_{\text{OSDA}} , the model p_{\theta}(x_{\text{comp}},x_{\text{cond}}|c_{\text{zeo}},c_{\text{OSDA}}) generates an ensemble of synthesis routes \{\tilde{x}_{\text{comp}}\} and \{\tilde{x}_{\text{cond}}\} via reverse diffusion (green arrow) that matches ground-truth synthesis routes \{x_{\text{comp}}\} , accurately capturing the aforementioned one-to-many structure-synthesis relationship with which previous methods struggle. Note: Only \{\tilde{x}_{\text{comp}}\} is shown. (c) Model architecture: Two separate encoders embed the zeolite material c_{\text{zeo}} and organic template c_{\text{OSDA}} before their latent representations are concatenated. The joint representation steers the generation process from noise ( x_{\text{comp},T},x_{\text{cond},T}\sim\mathcal{N}(0,\mathbf{I}) ) to realistic synthesis routes specific to the zeolite and OSDA via classifier-free guidance by learning \mu_{\theta}\left(x_{t},t,c_{\text{zeo}},c_{\text{OSDA}}\right)=\frac{1}{\sqrt{\alpha_{t}}}\left(x_{t}-\frac{\beta_{t}}{\sqrt{1-\bar{\alpha}_{t}}}\epsilon_{\theta}\left(x_{t},t,c_{\text{zeo}},c_{\text{OSDA}}\right)\right) where \epsilon_{\theta}\left(\cdot\right) is a conditional U-Net trained to predict the noise at time t (Section 4.2 ). Note that x_{\text{comp}} and x_{\text{cond}} are jointly noised and denoised. (d) PCA of learned representations of zeolites and OSDAs. The colorbar refers to a specific physical property (as shown in each title). 2 Results 2.1 DiffSyn framework for materials synthesis planning 2.1.1 Chemically guided diffusion model DiffSyn leverages a chemically guided diffusion model for predicting materials synthesis routes with the target zeolite structure c_{\text{zeo}} and the organic structure-directing agent (OSDA) c_{\text{OSDA}} as inputs (Fig. 1 b). An OSDA is an organic molecule that templates the zeolite’s pores (Fig. A1 e), directing the synthesis toward the formation of a specific structure. Prior work has shown that the optimal OSDA to synthesize a given structure can be identified from atomistic simulations lewis1995predicting ; zones1996guest ; jensen2021discovering ; schwalbe2021priori

Read the complete text