Automating alloy design and discovery with physics-aware multimodal multiagent AI

Proceedings of the National Academy of Sciences, Journal Year: 2025, Volume and Issue: 122(4)

Published: Jan. 24, 2025

The design of new alloys is a multiscale problem that requires holistic approach involves retrieving relevant knowledge, applying advanced computational methods, conducting experimental validations, and analyzing the results, process typically slow reserved for human experts. Machine learning can help accelerate this process, instance, through use deep surrogate models connect structural chemical features to material properties, or vice versa. However, existing data-driven often target specific objectives, offering limited flexibility integrate out-of-domain knowledge cannot adapt new, unforeseen challenges. Here, we overcome these limitations by leveraging distinct capabilities multiple AI agents collaborate autonomously within dynamic environment solve complex materials tasks. proposed physics-aware generative platform, AtomAgents, synergizes intelligence large language (LLMs) collaboration among with expertise in various domains, including retrieval, multimodal data integration, physics-based simulations, comprehensive results analysis across modalities. concerted effort multiagent system allows addressing problems, as demonstrated examples include designing metallic enhanced properties compared their pure counterparts. Our enable accurate prediction key characteristics highlight crucial role solid solution alloying steer development alloys. framework enhances efficiency multiobjective tasks opens avenues fields such biomedical engineering, renewable energy, environmental sustainability.

Language: Английский

---

OpenFOAMGPT: A retrieval-augmented large language model (LLM) agent for OpenFOAM-based computational fluid dynamics

Physics of Fluids, Journal Year: 2025, Volume and Issue: 37(3)

Published: March 1, 2025

This work presents a large language model (LLM)-based agent OpenFOAMGPT tailored for OpenFOAM-centric computational fluid dynamics (CFD) simulations, leveraging two foundation models from OpenAI: the GPT-4o (GPT means Generative Pre-trained Transformer) and chain-of-thought–enabled o1 preview model. Both agents demonstrate success across multiple tasks. While price of token with is six times as that GPT-4o, it consistently exhibits superior performance in handling complex tasks, zero-shot/few-shot case setup to boundary condition modifications, zero-shot turbulence adjustments, code translation. Through an iterative correction loop, efficiently addressed single-phase multiphase flow, heat transfer, Reynolds-averaged Navier–Stokes modeling, eddy simulation, other engineering scenarios, often converging limited number iterations at low costs. To embed domain-specific knowledge, we employed retrieval-augmented generation pipeline, demonstrating how preexisting simulation setups can further specialize subdomains such energy aerospace. Despite great agent, human oversight remains crucial ensuring accuracy adapting shifting contexts. Fluctuations over time suggest need monitoring mission-critical applications. Although our demonstrations focus on OpenFOAM, adaptable nature this framework opens door developing LLM-driven into wide range solvers codes. By streamlining CFD approach has potential accelerate both fundamental research industrial advancements.

Language: Английский

---

Accelerating Scientific Discovery with Generative Knowledge Extraction, Graph-Based Representation, and Multimodal Intelligent Graph Reasoning

Machine Learning Science and Technology, Journal Year: 2024, Volume and Issue: 5(3), P. 035083 - 035083

Published: Aug. 21, 2024

Abstract Leveraging generative Artificial Intelligence (AI), we have transformed a dataset comprising 1000 scientific papers focused on biological materials into comprehensive ontological knowledge graph. Through an in-depth structural analysis of this graph, calculated node degrees, identified communities along with their connectivities, and evaluated clustering coefficients betweenness centrality pivotal nodes, uncovering fascinating architectures. We find that the graph has inherently scale-free nature, shows high level connectedness, can be used as rich source for downstream reasoning by taking advantage transitive isomorphic properties to reveal insights unprecedented interdisciplinary relationships answer queries, identify gaps in knowledge, propose never-before-seen material designs, predict behaviors. Using large language embedding model compute deep representations use combinatorial similarity ranking develop path sampling strategy allows us link dissimilar concepts previously not been related. One comparison revealed detailed parallels between Beethoven’s 9th Symphony, highlighting shared patterns complexity through mapping. In another example, algorithm proposed innovative hierarchical mycelium-based composite based integrating principles extracted from Kandinsky’s ‘Composition VII’ painting. The resulting integrates set include balance chaos order, adjustable porosity, mechanical strength, complex patterned chemical functionalization. uncover other isomorphisms across science, technology art, revealing nuanced ontology immanence context-dependent heterarchical interplay constituents. Because our method transcends established disciplinary boundaries diverse data modalities (graphs, images, text, numerical data, etc), graph-based AI achieves far higher degree novelty, explorative capacity, technical detail, than conventional approaches establishes widely useful framework innovation hidden connections.

Language: Английский

---

Towards an automated workflow in materials science for combining multi-modal simulation and experimental information using data mining and large language models

Materials Today Communications, Journal Year: 2025, Volume and Issue: unknown, P. 112186 - 112186

Published: March 1, 2025

Language: Английский

---

Learning the rules of peptide self-assembly through data mining with large language models

Science Advances, Journal Year: 2025, Volume and Issue: 11(13)

Published: March 26, 2025

Peptides are ubiquitous and important biomolecules that self-assemble into diverse structures. Although extensive research has explored the effects of chemical composition exterior conditions on self-assembly, a systematic study consolidating these data to uncover global rules is lacking. In this work, we curate peptide assembly database through combination manual processing by human experts large language model–assisted literature mining. As result, collect over 1000 experimental entries with information about sequence, conditions, corresponding self-assembly phases. Using data, machine learning models developed, demonstrating excellent accuracy (>80%) in phase classification. Moreover, fine-tune GPT model for mining developed dataset, which markedly outperforms pretrained extracting from academic publications. This workflow can improve efficiency when exploring potential self-assembling candidates, guiding while also deepening our understanding governing mechanisms.

Language: Английский

---

Fine-tuning large language models for domain adaptation: exploration of training strategies, scaling, model merging and synergistic capabilities

npj Computational Materials, Journal Year: 2025, Volume and Issue: 11(1)

Published: March 28, 2025

Language: Английский

---

Graph-aware isomorphic attention for adaptive dynamics in transformers

APL Machine Learning, Journal Year: 2025, Volume and Issue: 3(2)

Published: April 22, 2025

We present an approach for modifying transformer architectures by integrating graph-aware relational reasoning into the attention mechanism, merging concepts from graph neural networks and language modeling. Building on inherent connection between theory, we reformulate transformer’s mechanism as a operation propose isomorphic attention. This method leverages advanced modeling strategies, including Graph Isomorphism Networks (GINs), to enrich representation of structures. Our improves model’s ability capture complex dependencies generalize across tasks, evidenced reduced generalization gap improved learning performance. expand concept introduce sparse-GIN-attention, fine-tuning that enhances adaptability pre-trained foundational models with minimal computational overhead, endowing them capabilities. show sparse-GIN-attention framework compositional principles category theory align sparsified structures while hierarchical bridges local interactions global task objectives diverse domains. results demonstrate mechanisms outperform traditional in both training efficiency validation These insights bridge uncover latent graph-like within mechanisms, offering new lens through which transformers can be optimized. By evolving GIN models, reveal their implicit capacity graph-level profound implications model development applications bioinformatics, materials science, modeling, beyond, setting stage interpretable generalizable strategies.

Language: Английский

---

Biomaterials in cellular agriculture and plant-based foods for the future

Nature Reviews Materials, Journal Year: 2025, Volume and Issue: unknown

Published: May 7, 2025

Language: Английский

---

Autonomous Self‐Evolving Research on Biomedical Data: The DREAM Paradigm

Advanced Science, Journal Year: 2025, Volume and Issue: unknown

Published: May 8, 2025

Abstract In contemporary biomedical research, the efficiency of data‐driven methodologies is constrained by large data volumes, complexity tool selection, and limited human resources. To address these challenges, a Data‐dRiven self‐Evolving Autonomous systeM (DREAM) developed as first fully autonomous research system capable independently conducting scientific investigations without intervention. DREAM autonomously formulates evolves questions, configures computational environments, performs result evaluation validation. Unlike existing semi‐autonomous systems, operates manual intervention validated in real‐world scenarios. It exceeds average performance top scientists question generation, achieves higher success rate environment configuration than experienced researchers, uncovers novel findings. context Framingham Heart Study, it demonstrated an that over 10 000 times greater scientists. As autonomous, self‐evolving system, offers robust efficient solution for accelerating discovery advancing other disciplines.

Language: Английский

---

Exploring structure–property relationships in sparse data environments using mixture-of-experts models

MRS Bulletin, Journal Year: 2024, Volume and Issue: unknown

Published: Nov. 13, 2024

Language: Английский

---