A Foundation Model for Material Fracture Prediction
Agnese Marcato, Aleksandra Pachalieva, Ryley G. Hill, Kai Gao, Xiaoyu Wang, Esteban Rougier, Zhou Lei, Vinamra Agrawal, Janel Chua, Qinjun Kang, Jeffrey D. Hyman, Abigail Hunter, Nathan DeBardeleben, Earl Lawrence, Hari Viswanathan, Daniel O’Malley, Javier E. Santos · 2025
Read A Foundation Model for Material Fracture Prediction by Agnese Marcato, Aleksandra Pachalieva, Ryley G. Hill, Kai Gao, Xiaoyu Wang, Esteban Rougier, Zhou Lei, Vinamra Agrawal, Janel Chua, Qinjun Kang, Jeffrey D. Hyman, Abigail Hunter, Nathan DeBardeleben, Earl Lawrence, Hari Viswanathan, Daniel O’Malley, Javier E. Santos (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-2507-23077.txt.
Introduction
Agnese Marcato and co-authors train one model to predict how materials break. Their 2025 paper presents a foundation model for fracture, a transformer that learns across simulators, materials, and loading conditions instead of one narrow setup at a time.
Its training spans explosives, steel, aluminum, shale, and tungsten, on both neat and messy meshes. Text descriptions of materials, boundaries, and solver settings enter through language model embeddings, so new scenarios need no redesign. Fine-tuned lightly, the model estimates time to failure, follows crack growth, and adapts to a new simulation method. It even handled titanium and concrete, materials it never trained on, from as little as one example. The authors offer it as a shared base for failure prediction.
Catalog record
| CREATOR | Agnese Marcato, Aleksandra Pachalieva, Ryley G. Hill, Kai Gao, Xiaoyu Wang, Esteban Rougier, Zhou Lei, Vinamra Agrawal, Janel Chua, Qinjun Kang, Jeffrey D. Hyman, Abigail Hunter, Nathan DeBardeleben, Earl Lawrence, Hari Viswanathan, Daniel O’Malley, Javier E. Santos |
|---|---|
| YEAR | 2025 |
| COLLECTION | PAPERS |
| FORMAT | TXT |
| FILE SIZE | 57 KB |
| VERSION | 1.0 |
| LICENSE | http://creativecommons.org/licenses/by/4.0/ |
| RIGHTS | CC BY |
| SOURCE | arXiv |
Source: https://arxiv.org/abs/2507-23077
Browse the full library · About the Tower of Babel
Full text
The opening of the work, from the library's hosted text:
A Foundation Model for Material Fracture Prediction Agnese Marcato, Aleksandra Pachalieva, Ryley G. Hill, Kai Gao, Xiaoyu Wang, Esteban Rougier, Zhou Lei, Vinamra Agrawal, Janel Chua, Qinjun Kang, Jeffrey D. Hyman, Abigail Hunter, Nathan DeBardeleben, Earl Lawrence, Hari Viswanathan, Daniel O’Malley, Javier E. Santos Abstract Accurately predicting when and how materials fail is critical to the design of safe and reliable structures, mechanical systems, and engineered components that operate under stress. Yet, fracture behavior remains difficult to model across the diversity of materials, geometries, and loading conditions encountered in real-world applications. While machine learning (ML) methods have shown promise, most models are trained on narrow datasets, lack robustness, and struggle to generalize to new regimes. Meanwhile, physics-based simulators offer high-fidelity predictions but are fragmented across specialized numerical methods and require substantial high-performance computing resources to explore the input space. To address these limitations, we present a data-driven foundation model for fracture prediction, a transformer-based architecture that operates across simulators, a wide range of materials (including plastic-bonded explosives, steel, aluminum, shale, and tungsten) and diverse loading conditions. The model supports both structured and unstructured meshes, combining them with large language model embeddings of textual input decks that specify material properties, boundary conditions, and solver settings. This multimodal input design enables flexible adaptation across diverse simulation scenarios without requiring changes to the model architecture. The trained model can be fine-tuned with minimal data on diverse downstream tasks. These include time-to-failure estimation, modeling the temporal evolution of fractures, and adapting to combined finite-discrete element method simulations. The model also generalizes to previously unseen materials such as titanium and concrete, requiring as few as a single sample, dramatically reducing data requirements compared to standard ML approaches. Our results demonstrate that fracture prediction can be unified under a single model architecture, offering a scalable, extensible alternative to simulator-specific workflows. keywords Material failure, Foundation model, Transformer, Multimodal architecture, LLM integration Fractures govern the limits of performance and safety in both engineered and natural systems. From structural collapse and brittle failure in critical infrastructure Qu et al. 2024 to pipeline rupture Niazi et al. 2021 , the consequences of material failure span a wide range of applications. In contexts such as hydraulic fracturing and subsurface containment of hazardous materials Hyman et al. 2016 , understanding when and how geo-materials break is essential not only for mitigating risks such as fluid leakage and induced seismicity, but also for the design and optimization of subsurface engineering systems Launey and Ritchie 2009 ; Kalthoff 1986 . Despite decades of research, reliably predicting fracture behavior across materials and loading conditions remains a fundamental challenge. This has led to the development of numerous specialized numerical simulators, each tailored to address specific modeling hurdles. Mesh-free and particle methods, such as peridynamics and the material point method, handle extreme deformations without mesh entanglement Silling 2000 ; Sulsky et al. 1995 . Interface-enriched finite elements like XFEM embed discontinuities directly in the solution space, capturing sharp cracks without remeshing Belytschko and Black 1999 . Phase-field models regularize cracks as diffusive fields, allowing for branching and merging on fixed grids Bourdin et al. 2000 , while cohesive-zone and adaptive meshing techniques explicitly evolve mesh topology to follow complex crack paths Ortiz and Pandolfi 1999 . Although effective within their respective regimes, these methods often suffer from limitations such as high computational cost, the need for a priori knowledge of crack paths, missing material models, and reliance on expert parameter tuning for physically consistent results. Practitioners often face uncertainty in selecting the appropriate fracture simulator. Each solver family has been developed, benchmarked, and validated within its regime, but lacks interoperability; transferring between them requires re-encoding meshes, constitutive laws, and boundary conditions. We address this gap with a simulation-level foundation model that learns a shared latent representation across diverse fracture solvers. Using inexpensive rule-based surrogate simulations for pretraining, the model learns generic topological operations such as crack growth and coalescence, then adapts to high-fidelity phase-field and finite-discrete element data via a multi-fidelity curriculum. This strategy reduces data demands and enables rapid cross-solver inference for example, replacing XFEM screening sweeps with single-GPU predictions. Similar multi-fidelity approaches are already used in aerospace, where Gaussian-process surrogates and neural networks combine RANS and LES data for efficient aerodynamic optimization Scoggins et al. 2023 ; Tao et al. 2024 . Recently, machine learning (ML) approaches have emerged as a tool that offers the potential to learn predictive models directly from data. Among these, foundation models Bommasani et al. 2021 are large machine learning models trained on broad and diverse datasets, enabling them to learn general-purpose representations that can be adapted to a wide range of downstream tasks. These models are typically pretrained and fine-tuned for specific applications, leveraging their ability to transfer knowledge across domains. Their defining features include scalability, versatility, and strong performance with minimal task-specific data. Foundation models provide a framework for capturing patterns across domains and efficiently adapting to new tasks by exploiting knowledge acquired during pretraining Brown et al. 2020 . Applications in weather forecasting Bodnar et al. 2025 ; Palmer et al. 2008 , protein structure prediction Jumper et al. 2021 ; Stokes et al. 2020 , and learning interatomic potentials for molecular dynamics Rupp et al. 2012 ; Ahmad et al. 2022 ; Batatia et al. 2023 demonstrate that training on diverse datasets can yield high-performing systems that generalize across tasks and offer a scalable, efficient alternative to traditional simulation or data collection pipelines by aggregating available information and delivering predictions orders of magnitude faster than conventional methods. Foundation models are increasingly impacting materials science Pyzer-Knapp et al. 2025 ; Pyzer-Knapp et al. 2022 : enabling automatic data extraction from the literature Gupta et al. 2022 ; Rajan et al. 2023 , prediction of molecular properties and structures Takeda et al. 2023 ; Chithrananda et al. 2020 ; Ross et al. 2022 ; Bagal et al. 2021 , and support for material synthesis Chang and Ye 2024 ; Kim et al. 2020 . Specifically for fracture propagation in materials, ML methods have been applied to predict crack evolution Lew et al. 2021 ; Perera et al. 2022 ; Hsu et al. 2020 ; Zhang and Wei 2022 ; Perera and Agrawal 2023 ; Chen et al. 2024 , crack coalescence Moore et al. 2018 ; Hunter et al. 2019 , stress distribution Wang et al. 2021 ; Perera and Agrawal 2023 ; Perera et al. 2022 , fracture toughness Liu et al. 2020 ; Daghigh et al. 2020 , fatigue strength Yan et al. 2020 ; Frie et al. 2024 ; Wang et al. 2024 , time to failure for materials Hunter et al. 2019 ; Biswas et al. 2020 , and accelerating workflows Liu and Misra 2022 ; Perry et al. 2022 ; Thi et al. 2025 . However, most ML models for fracture propagation are typically trained on narrow datasets from a single simulator, under fixed material and geometric conditions. As a result, they struggle to generalize across simulation settings, limiting their practical utility and reproducibility. Transfer learning has been shown to alleviate some of these concerns Yosinski et al. 2014 ; Zhou and Yu 2024 ; Perera and Agrawal 2023 , but the required