Skill-Based Autonomous Agents for Material Creep Database Construction

TL;DR

This work presents an autonomous, skill-based agent framework that leverages a foundation LLM to extract, validate, and organize creep-mechanics data from unstructured literature. Through a five-stage pipeline—literature collection, automated screening, multi-modal extraction, physics-informed validation, and structured storage—the system builds a physically self-consistent creep database from 243 publications, achieving over 90% verified extraction accuracy. A modular Cognitive Architecture with explicit constraints and cross-modal alignment ensures data fidelity, provenance via DOIs, and an accessible web interface for querying. The results demonstrate broad material coverage, high-fidelity curve-digitization, and near-perfect cross-modal agreement ($R^2=0.9999$) between data and constitutive models, highlighting the framework's potential to generalize to other domains and accelerate AI-assisted scientific discovery.

Abstract

The advancement of data-driven materials science is currently constrained by a fundamental bottleneck: the vast majority of historical experimental data remains locked within the unstructured text and rasterized figures of legacy scientific literature. Manual curation of this knowledge is prohibitively labor-intensive and prone to human error. To address this challenge, we introduce an autonomous, agent-based framework powered by Large Language Models (LLMs) designed to excavate high-fidelity datasets from scientific PDFs without human intervention. By deploying a modular "skill-based" architecture, the agent orchestrates complex cognitive tasks - including semantic filtering, multi-modal information extraction, and physics-informed validation. We demonstrate the efficacy of this framework by constructing a physically self-consistent database for material creep mechanics, a domain characterized by complex graphical trajectories and heterogeneous constitutive models. Applying the pipeline to 243 publications, the agent achieved a verified extraction success rate exceeding 90% for graphical data digitization. Crucially, we introduce a cross-modal verification protocol, demonstrating that the agent can autonomously align visually extracted data points with textually extracted constitutive parameters ($R^2 > 0.99$), ensuring the physical self-consistency of the database. This work not only provides a critical resource for investigating time-dependent deformation across diverse material systems but also establishes a scalable paradigm for autonomous knowledge acquisition, paving the way for the next generation of self-driving laboratories.

Figures (7)

• Figure 1: Architecture of the automated literature mining pipeline. The workflow proceeds in five distinct stages: (1) Literature Collection: Aggregation of source PDFs from scientific repositories; (2) Automated Screening: An LLM-based filter analyzes document content to identify relevant publications containing creep constitutive models; (3) Information Extraction: A multi-modal extraction engine parses both unstructured text (for material metadata and formulas) and graphical plots (for experimental creep trajectories); (4) Formula Validation: A logic-based validation module checks the mathematical integrity and physical plausibility of the extracted expressions; (5) Structured Storage & Interface: Validated data is serialized into a relational database, accessible via a user-friendly web interface for searching, filtering, and exporting.
• Figure 2: Interactive front-end interface of the AI constructed Creep Database. The web-based platform acts as a bridge between the structured SQL backend and end users, facilitating efficient data retrieval. Key features include: (1) Granular Filtering (Left Panel): Users can query the database using multidimensional constraints, such as material composition, temperature ranges, and stress levels; (2) Dynamic Visualization (Right Panel): Selected datasets are instantly rendered into strain time plots, allowing for rapid visual comparison of creep behaviors across different experimental conditions; (3) Data Export: Validated data points and constitutive parameters can be exported in standardized formats for downstream machine learning applications.
• Figure 3: Demonstration of selective curve extraction from complex multi-variable plots. (a) Original source image (Material: X46Cr13, T=600$^\circ$C) containing multiple creep trajectories under different stress levelsma10040388. (b) The digitized result reconstructed from the agent's output. The agent successfully disentangled the visual information, accurately isolating the specific curve corresponding to the target condition ($\sigma = 31.6$ MPa) while ignoring the extraneous data series (52.7 MPa). Note that the high fidelity of the extracted data points preserves the underlying creep trend.
• Figure 4: Diversity of material composition within the constructed creep database. The sector chart illustrates the distribution of material categories across the 353 extracted creep curves. The dataset demonstrates broad domain coverage, encompassing critical industrial materials (e.g., Nickel-based alloys, 19.0%; Steel/Iron, 12.5%) as well as non-metallic systems (e.g., Polymers, 8.5%; Ice/Glaciers, 4.0%). This diversity confirms the agent's semantic ability to identify relevant "creep" behaviors across heterogeneous material science sub-domains.
• Figure 5: Statistical distribution of experimental conditions extracted from the literature. (a) Frequency distribution of experimental temperatures ($^\circ$C), spanning from cryogenic environments to ultra-high temperature regimes ($>1000^\circ$C). (b) Frequency distribution of applied stress levels (MPa). The broad range of operating conditions indicates that the database captures a comprehensive snapshot of material performance boundaries documented in historical literature.