8 results
Now showing1 - 8 of 8
- Some of the metrics are blocked by yourconsent settings
Item type:Publication, Access status: with files , LLM-driven design of physics-constrained constitutive models: Two agents are better than one(Springer, 2026-10-01); ; ; ; Developing constitutive models that capture how materials deform under load traditionally requires years of specialized expertise in continuum mechanics, machine learning, and scientific programming. Large language models (LLMs) have recently been shown to lower this barrier by generating constitutive models on demand, but existing single-agent pipelines lack systematic checks that the resulting models respect fundamental physical laws. To close this gap, we introduce the first multi-agent LLM-driven approach for constitutive model generation: a Creator agent proposes a model tailored to the data, while an Inspector agent critically audits each proposal against nine physical constraints and returns it for refinement whenever a violation is detected. We demonstrate this concept with constitutive artificial neural networks (CANNs) and benchmark it on brain tissue and rubber as isotropic materials, and on porcine skin tissue as a transversely isotropic material with a preferred fiber direction, using two different LLM backbones (Claude Opus 4.7 and Kimi K2.5). Whether a generated model satisfies the physical constraints is assessed numerically, by probing each constraint across a broad sample of deformation states, rotations, and perturbation directions. Adding the Inspector raises the share of exported models that pass all these checks from 90% to 95% for Opus and from 47% to 60% for Kimi. In addition, the generated models are on par with or even surpass expert-designed models in accuracy, extrapolate reliably beyond the training data, and generalize remarkably well to unseen loading paths. Separating generation from inspection thus turns LLM-driven constitutive modeling into a substantially more trustworthy process. The paradigm is deliberately technique-agnostic and scales automatically with advances in LLM capability, opening a promising path toward automated, physics-aware model discovery.Publicationtype: Journal Article7 28 - Some of the metrics are blocked by yourconsent settings
Item type:Publication, Access status: with files , Different alloy, different story? Impact of transfer learning on predictive models for in silico screening of Mg dissolution modulators(Elsevier, 2026-09-26); ; ; ; Ideally, datasets for training machine learning models consist of thousands of versatile data samples — a scale that is seldom met in materials research to date. Transfer learning (TL) is widely used to mitigate data scarcity in available training sets by leveraging knowledge from related domains, since models pretrained on auxiliary datasets may learn structure–property relationships that are transferable from one to another. In this work, the potential of TL for predicting the performance of small organic molecule dissolution modulators is probed using corrosion inhibition responses of eleven Mg-based materials taken from the previously published ExCorr database. A significant improvement in prediction accuracy is observed when transfer learning is applied. However, for the dataset configuration studied here, the apparent cross-alloy TL gains are shown not to arise from transferable structure–activity relationship (SAR) knowledge but rather from a learned alloy-context signal embedded in the data. Seven Mg-based materials—six commercial alloys plus the high-purity variant HPMg51—form a shared-SAR cluster with strongly correlated inhibitor rankings, such that molecules labelled on one material are readily predictable for others within the cluster. Within this shared-SAR cluster, extensive overlap reduces the effective training pool from approximately 900 records to only around 180 unique molecules, placing the task firmly in a small-data regime. In this setting, an in-context large language model (Gemini 3.1 Pro) is the only approach that successfully learns meaningful SAR, with its pretrained knowledge providing the transfer that the explicit source dataset fails to deliver. Seven Mg-based materials in the ExCorr database form a shared-structure–activity relationship (SAR) cluster (pairwise Spearman r=0.67–0.92). Under molecule-disjoint filtering the cross-alloy TL gain on ExCorr splits into a null inhibitor-SAR contribution and a dominant alloy-context contribution. Cross-alloy overlap shrinks ExCorr’s training pool from ≈900records to ≈180unique molecules, only an in-context LLM (Gemini 3.1 Pro) stays predictive.Publicationtype: Journal Article7 5 - Some of the metrics are blocked by yourconsent settings
Item type:Publication, Access status: with files , LLMs as designers of physics-constrained neural networks for constitutive modeling: a demonstration on hyperelastic solids(Springer, 2026-06-12); ; ; ; Large language model (LLM)-based frameworks extend beyond agents; they also enable the on-demand creation of specialized scientific and engineering tools. We demonstrate this concept in the field of solid mechanics. There, so-called constitutive models describe the relationship between body deformation and mechanical stress. They are essential for both the scientific understanding and industrial application of materials. However, even recent data-driven methods of constitutive modeling, such as constitutive artificial neural networks (CANNs), still require substantial expert knowledge and human labor. We present a framework in which an LLM generates a CANN on demand, tailored to a given material class and dataset provided by the user. The framework covers LLM-based architecture selection, integration of physical constraints, and complete implementation. Evaluation on three benchmark problems demonstrates that LLM-generated CANNs achieve accuracy and generalization comparable to or greater than manually engineered counterparts, while substantially reducing the expertise required for constitutive modeling.Publicationtype: Journal Article33 25 - Some of the metrics are blocked by yourconsent settings
Item type:Publication, Access status: without files , Concurrently coupling particle and continuum simulations to study block copolymer membrane fabrication(2026-06-02); ; ; ; We present a concurrent multiscale simulation framework for membrane fabrication from a block copolymer solution via self-assembly and nonsolvent-induced phase separation, combining simulations of a soft, coarse-grained particle model with the continuum Uneyama-Doi model. The computationally intensive particle model provides a molecularly resolved description of micro- and macrophase separation, including thermal fluctuations, while the computationally efficient continuum model captures the process-driven self-assembly of nonequilibrium membrane morphologies on large length and time scales. Central to the approach is a machine learning-guided adaptive coupling strategy, implemented through a coordinator library, which predicts the evolving spatiotemporal subdomain where high-fidelity particle simulations are required and dynamically allocates computational resources accordingly. The two complementary models are concurrently coupled through a consistent exchange of the solvents' fluxes, enabling the treatment of spatially inhomogeneous, multicomponent systems with diffusive transport, micro- and macrophase separation, and vitrification. This adaptive strategy enables predictive simulations of membrane formation on experimentally and technologically relevant length and time scales, reaching micrometers and minutes. As an application, we examine the influence of polymer concentration in the initial casting solution. The framework is general and extensible, providing a computational tool for investigating nonequilibrium structure formation in complex multicomponent soft-matter systems.Publicationtype: Journal Article17 - Some of the metrics are blocked by yourconsent settings
Item type:Publication, Access status: with files , Adversarial robustness of LLM-based multi-agent systems for engineering problems(Frontiers Media S.A., 2026-05-13); ; ; ; Large language models (LLMs) are increasingly deployed in multi-agent systems (MAS), including for solving engineering problems. Unlike purely linguistic tasks, engineering workflows demand formal rigor and numerical accuracy, meaning that adversarial perturbations can cause not just degraded performance but systematically incorrect or unsafe results. In this work, we present one of the first systematic studies of adversarial robustness of LLM-based MAS in engineering contexts. Using representative problems—including pipe pressure loss (Darcy-Weisbach), beam deflection, mathematical modeling, and graph traversal—we investigate how misleading agents affect collaborative reasoning and quantify error propagation under controlled adversarial influence. Our results show that adversarial vulnerabilities in engineering differ from those observed in generic MAS evaluations in important aspects: system robustness is sensitive to task type, the subtlety of injected errors, and communication order among agents. In particular, engineering tasks with higher structural complexity or easily confusable numerical variations are especially prone to adversarial influence. We further identify design choices, such as prompt framing, agent role assignment, and discussion order, that significantly improve resilience. These findings highlight the need for domain-specific evaluation of adversarial robustness and provide actionable insights for designing MAS that are trustworthy and safe in engineering applications.Publicationtype: Journal Article45 43 - Some of the metrics are blocked by yourconsent settings
Item type:Publication, Access status: with files , Functional partitioning through competitive learning(Frontiers Media S.A., 2025-11-05); ; ; ; Datasets often incorporate various functional patterns related to different aspects or regimes, which are typically not equally present throughout the dataset. We propose a novel partitioning algorithm that utilizes competition between models to detect and separate these functional patterns. This competition is induced by multiple models iteratively submitting their predictions for the dataset, with the best prediction for each data point being rewarded with training on that data point. This reward mechanism amplifies each model's strengths and encourages specialization in different patterns. The specializations can then be translated into a partitioning scheme. We validate our concept with datasets with clearly distinct functional patterns, such as mechanical stress and strain data in a porous structure. Our partitioning algorithm produces valuable insights into the datasets' structure, which can serve various further applications. As a demonstration of one exemplary usage, we set up modular models consisting of multiple expert models, each learning a single partition, and compare their performance on more than twenty popular regression problems with single models learning all partitions simultaneously. Our results show significant improvements, with up to 56% loss reduction, confirming our algorithm's utility.Publicationtype: Journal Article19 10 - Some of the metrics are blocked by yourconsent settings
Item type:Publication, Access status: with files , Large language models predicting the corrosion inhibition efficiency of magnesium dissolution modulators(Elsevier, 2025-06-20); ; ; ; Large language models (LLMs), such as GPT-4o, have shown promise in solving everyday tasks and addressing fundamental scientific challenges by leveraging extensive pre-trained knowledge. In this study, we investigate their potential to predict the efficiency of various organic compounds in inhibiting the corrosion of the magnesium alloy ZE41. Traditional approaches, such as Multilayer Perceptrons (MLPs), rely on non-contextual data, often necessitating large datasets and substantial effort per sample to achieve accurate predictions. These methods particularly struggle with small datasets as their training data and domain of applicability is limited to a small area of the available chemical space. LLMs can contextualize and interpret limited data points by drawing on their vast knowledge, including the chemical properties of molecules and their influence on corrosion processes in other materials (e.g. iron and aluminium). By prompting the model with a small dataset, LLMs can provide meaningful predictions without the need for extensive training. Our study demonstrates that LLMs can predict corrosion inhibition outcomes and outperform classical approaches, such as MLPs, having access to the identical number of training samples.Publicationtype: Journal Article34 23 - Some of the metrics are blocked by yourconsent settings
Item type:Publication, Access status: with files , Constitutive scientific generative agent (CSGA): leveraging large language models for automated constitutive model discovery(Springer, 2025-05-29); ; ; Data-driven approaches for constitutive modeling enable rapid, automated generation of models that predict a material’s mechanical response under load. Integrating theoretical knowledge into these approaches, which are then called grey-box approaches, can improve sample efficiency, extrapolation capability, and interpretability, albeit typically at the cost of experts required to use them. Recently, general-purpose large language model (LLM)-based scientific discovery methods have emerged as user-friendly approaches to scientific discovery. In this work, we compare two representatives of these paradigms: highly specialized constitutive artificial neural networks (CANNs) and the general LLM-based scientific generative agent (SGA) to evaluate current LLM capabilities in constitutive modeling. In addition, we introduce the constitutive scientific generative agent (CSGA) to combine both approaches’ strengths by enriching the SGA’s prompts with domain-specific data and materials theory. We compare CANN, SGA, and CSGA on three benchmark problems by assessing their accuracy in predicting stress responses under prescribed strain conditions. While our results show that CANNs remain the most accurate approach overall, the CSGA significantly outperforms the SGA and demonstrates the promise of specialized LLM-based methods for constitutive modeling. Moreover, the CSGA’s intuitive plain text interface and the full interpretability of the generated constitutive models make it a practical, accessible complement to existing approaches.Publicationtype: Journal Article25 23