Kévin Alhada-Lahbabi
1 · 2025
Read Kévin Alhada-Lahbabi by 1 (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-2501-05312.txt.
Introduction
Kevin Alhada-Lahbabi and co-authors link build to behavior in a memory film material in a 2025 paper. The film, a blend of two oxides, suits tiny electronics, but how its grain pattern shapes its electric response stayed murky.
The team pair physics based film growth runs with a learning model that squeezes response loops into a small code. Varying grain size, polar share, and grain tilt feeds the data set. The code teases apart traits that usually move together. Flipping the tool around lets them design loop shapes to order, tuning switch force and stored charge for a target device.
Catalog record
| CREATOR | 1 |
|---|---|
| YEAR | 2025 |
| COLLECTION | PAPERS |
| FORMAT | TXT |
| FILE SIZE | 39 KB |
| VERSION | 1.0 |
| LICENSE | http://creativecommons.org/licenses/by/4.0/ |
| RIGHTS | CC BY |
| SOURCE | arXiv |
Source: https://arxiv.org/abs/2501-05312
Browse the full library · About the Tower of Babel
Full text
The opening of the work, from the library's hosted text:
Kévin Alhada-Lahbabi 1 Abstract While Hf 0.5 Zr 0.5 O 2 (HZO) thin films hold significant promise for modern nanoelectronic devices, a comprehensive understanding of the interplay between their polycrystalline structure and electrical properties remains elusive. Here, we present a novel framework combining phase-field (PF) modeling with Variational Autoencoders (VAEs) to uncover structure-property correlations in polycrystalline HZO. Leveraging PF simulations, we constructed a high-fidelity dataset of P P - V V loops by systematically varying critical material parameters, including grain size, polar grain fraction, and crystalline orientation. The VAEs effectively encoded hysteresis loops into a low-dimensional latent space, capturing electrical properties while disentangling complex material parameters interdependencies. We further demonstrate a VAE-based inverse design approach to optimize P P - V V loop features, enabling the tailored design of device-specific key performance indicators (KPIs), including coercive field, remanent polarization, and loop area. The proposed approach offers a pathway to systematically explore and optimize the material design space for ferroelectric nanoelectronics. marginparsep has been altered. topmargin has been altered. marginparpush has been altered. The page layout violates the ICML style. Please do not change the page layout, or include packages like geometry, savetrees, or fullpage, which change it for you. We’re not able to reliably undo arbitrary changes to the style. Please remove the offending package(s), or layout-changing commands and try again. Unveiling structure-property correlations in ferroelectric Hf 0.5 Zr 0.5 O 2 films using variational autoencoders † † footnotetext: 1 . Correspondence to: [email protected], [email protected] <>. 1 Introduction Ferroelectric materials have been foundational in information and communication technologies for decades, with the first commercial FeRAMs (Ferroelectric Random-Access Memories) emerging in the 1990s. These perovskite-based devices demonstrated low-voltage, energy-efficient switching, consuming only 1-10 fJ/bit—up to three orders of magnitude lower than other memory technologies ( 1 ) . However, their scalability has been limited to the 130 nm technology node, and their chemical incompatibility with CMOS technology hindered further advancements ( 2 ) . In contrast, fluorite-structured ferroelectrics, such as Hf 0.5 Zr 0.5 O 2 (HZO), have emerged as promising candidates for next-generation devices due to their compatibility with CMOS ( 3 ) and their demonstrated ferroelectricity in ultra-thin films (below a thickness of 2 nm) ( 4 ) . Despite these advantages, HZO-based ferroelectric devices face critical challenges, including wake-up, fatigue, and imprint. Wake-up refers to the gradual improvement in polarization states during initial cycling, while fatigue describes the eventual degradation of these states, reducing the device’s endurance and reliability ( 5 ) . Imprint, characterized by the shift in polarization hysteresis loops due to charge trapping or internal electric fields, further exacerbates variability and impacts long-term stability. These phenomena are intricately linked to nanoscale structural features such as grain boundaries, phase heterogeneities, and local defects, resulting in inconsistent performance across devices even under identical processing conditions. To address these challenges, a multi-scale characterization framework is essential, bridging nanoscale material behavior with macroscopic device performance. Such an approach can correlate structural features like grain size, interfaces, and phase composition with device metrics such as endurance, retention, and coercive field. The complexity and non-linearity of these relationships pose significant challenges for conventional analytical methods. Recent advances in unsupervised machine learning, have demonstrated the potential to uncover latent physical mechanisms through variational autoencoders, offering transformative solutions for linking material behavior across scales ( 6 ; 7 ) . VAEs are particularly well-suited for handling high-dimensional, multi-scale data. They compress complex datasets into a low-dimensional latent space, uncovering hidden patterns and key variables driving material behavior ( 8 ; 9 ) . This allows for the identification of critical correlations between nanoscale properties and macroscopic performance, such as the impact of grain size distribution on fatigue resistance or the role of phase heterogeneity in wake-up behavior. Furthermore, VAEs enable predictive modeling by exploring the latent space to identify optimal configurations for desired device properties ( 10 ) , such as finding initial film parameters that maximize remnant polarization for increasing a device memory window and/or minimize the coercive fields to reduce the device energy consumption ( 11 ; 12 ) . The optimal film parameters for ferroelectric memory devices depend heavily on the specific device architecture and target application. For example, in Ferroelectric Field-Effect Transistors (FeFETs), minimizing leakage currents and optimizing the interface quality between the ferroelectric layer and the channel is critical for achieving high endurance and retention ( 13 ) . Similarly, for neuromorphic applications, controlling the polarization dynamics to attain stable, multi-level resistance states is crucial for mimicking synaptic behavior ( 14 ; 15 ) . By leveraging the latent space of VAEs, it becomes possible to tailor film properties to meet the specific performance requirements of each device type, enabling more targeted and efficient material optimization. High-throughput phase-field (PF) simulations have become fundamental for understanding ferroelectric materials, particularly Hf 0.5 Zr 0.5 O 2 (HZO)-based devices ( 16 ; 17 ; 18 ) . Recent studies have modeled ferroelectric hysteresis in HZO ( 19 ; 20 ; 21 ; 22 ; 23 ; 24 ; 25 ) and analyzed the effects of polycrystalline grain structures on electrical properties ( 26 ; 27 ) . Phase-field modeling has also been applied to investigate electric-field-induced phase transitions ( 28 ) , spacer layer impacts on domain dynamics ( 29 ) , and the role of oxygen vacancies in electrical behavior ( 30 ; 31 ; 32 ) . PF modeling thus enables the generation of high-fidelity datasets capturing polarization evolution and field-induced dynamics in HZO thin films. In this work, we first leverage PF modeling to simulate a wide range of material parameter distributions, such as variations in grain size, phase fraction, and defect distributions, to generate rich datasets for training the VAE. Next, we demonstrate how the VAE latent space disentangles the complex interplay of these material parameters and their impact on HZO’s electrical properties. Finally, we illustrate the inverse design of ferroelectric hysteresis properties using the VAE framework. This integration of PF simulation and VAE facilitates systematic exploration of the material design space, offering valuable guidance for experimental efforts to optimize functional properties. 2 Results 2.1 Material Parameter space In this study, our primary objective is to investigate how the interaction between key material parameters governing polycrystalline structures influences the electrical properties of HZO, utilizing PF modeling and VAE. Specifically, we aim to explore the impact of four critical parameters that dictate the crystalline arrangement: the average grain size ( D_{G} ), the proportion of orthombic polar to monoclinic non-polar grains ( \nu_{ortho} ), the preferential orientation of the crystallographic axis ( \mu_{\theta} ), and the dispersion of grain orientations relative to this axis ( \sigma_{\theta} ). In Figure 1 a, three examples of 10\,\text{nm}\times 128\,\text{nm} 2D polycrystalline structures, generated via Voronoi tessellation, are presented, each corresponding to a distinct combination of material parameters [D_{G},\nu_{\text{ortho}},\mu_{\theta},\sigma_{\theta}] . These structures exhibit various grain arrangements based on the average grain size D_{G} (here ranging from 10 nm to 20 nm), encompassing both equiaxed and columnar grains. The polarization vector field, initialized in a downward orientation, is represented by white arrows, illustrating the interplay between the fiber texture \mu_{\theta} and the grain orientation dispersion \sigma_{\theta} , as well as their effect on