AGAPI-Agents: An Open-Access Agentic AI Platform for Accelerated Materials Design on AtomGPT.org

AGAPI-Agents: An Open-Access Agentic AI Platform for Accelerated Materials Design on AtomGPT.org Jaehyung Lee1, Justin Ely2, Kent Zhang2, Akshaya Ajith1, Charles Rhys Campbell1,3, Kamal Choudhary1,2,∗ 1Department of Materials Science and Engineering, Johns Hopkins University, Baltimore, MD 21218, USA 2Department of Electrical and Computer Engineering, Johns Hopkins University, Baltimore, MD 21218, USA 3Department of Physics and Astronomy, West Virginia University, Morgantown, WV 26506, USA ∗Corresponding author: kchoudh2@jhu.edu December 16, 2025 Abstract Artificial intelligence is reshaping scientific discovery, yet its use in materials research remains lim- ited by fragmented computational ecosystems, reproducibility challenges, and dependence on commer- cial large language models (LLMs). Here we introduce AGAPI (AtomGPT.org API), an open-access agentic AI platform that integrates more than eight open-source LLMs with over twenty materials- science API endpoints, unifying databases, simulation tools, and machine-learning models through a common orchestration framework. AGAPI employs an Agent-Planner-Executor-Summarizer ar- chitecture that autonomously constructs and executes multi-step workflows spanning materials data retrieval, graph neural network property prediction, machine-learning force-field optimization, tight- binding calculations, diffraction analysis, and inverse design. We demonstrate AGAPI through end-to- end workflows, including heterostructure construction, powder X-ray diffraction analysis, and semi- conductor defect engineering requiring up to ten sequential operations. In addition, we evaluate AGAPI using 30+ example prompts as test cases and compare agentic predictions with and without tool access against experimental data. With more than 1,000 active users, AGAPI provides a scal- able and transparent foundation for reproducible, AI-accelerated materials discovery. AGAPI-Agents codebase is available at https://github.com/atomgptlab/agapi 1 Introduction The accelerating pace of scientific discovery increasingly demands the integration of heterogeneous compu- tational tools, expansive databases, and sophisticated machine learning models into coherent workflows[1, 2, 3, 4]. Large language models (LLMs) have emerged as promising orchestrators for such workflows, demonstrating remarkable capabilities in natural language understanding, multi-step reasoning, and code generation[5, 6, 7]. In materials science, LLMs show potential for tasks ranging from literature synthesis and experimental design to property prediction and inverse materials discovery[8, 9, 10, 11, 12, 13, 14, 15]. However, the deployment of LLMs in rigorous scientific contexts faces several fundamental challenges. First, hallucination remains a critical concern, where models generate outputs that are factually incorrect, physically inconsistent, or entirely fabricated, yet presented with high confidence[16, 17, 18]. In materials science, hallucinated predictions can suggest impossible crystal structures, erroneous property values, or chemically infeasible reactions, leading researchers down unproductive paths. Second, sycophancy bias causes LLMs to over-praise results rather than provide critical evaluation, creating false confidence in computational predictions[19]. Third, most existing agentic frameworks rely on commercial LLMs (e.g., GPT-4, Claude), introducing cost barriers, non-deterministic behavior across API versions, and potential intellectual property concerns when proprietary research data passes through external servers[20]. Current approaches to incorporate materials science knowledge in foundational LLMs fall into three categories: training LLMs from scratch on scientific corpora[21, 22], fine-tuning pre-trained models on domain-specific datasets[13], and developing agentic frameworks that augment LLMs with external 1 arXiv:2512.11935v1 [cs.AI] 12 Dec 2025 tools[11, 23, 24]. Training from scratch, while effective, requires computational resources beyond the reach of most research groups. Fine-tuning has shown success in materials science with models such as LLM- Prop[25], CrsystaLLM[26], CrystaLLM[15], AtomGPT[13], DiffractGPT[12], and MicroscopyGPT[14], but requires curated datasets for each application domain and lacks flexibility across multiple material classes. Agentic AI frameworks represent a complementary paradigm that leverages pre-trained LLMs as reasoning engines while connecting them to external tools, databases, and APIs through orchestrated workflows[27, 28, 29, 30]. This approach enables LLMs to plan, execute, and validate complex multi- step scientific tasks without costly retraining. Notable examples include Coscientist for autonomous chemical experimentation[24], ChemCrow for multi-tool chemistry workflows[23], AtomAgents for mate- rials simulations[31], AURA for NanoHub integration[32], LLamp for Materials Project integration[33], SciToolAgent[34], and ChatGPT Material Explorer[11], among oth

This content is AI-processed based on open access ArXiv data.