AtomGPT.org Ecosystem Overview

• AtomGPT.org Ecosystem is an open, modular platform that integrates transformer-driven models, physics-aware data representations, and open-access APIs for materials science and quantum research.
• It leverages both forward property prediction and generative inverse design using specialized model families like AtomGPT and RydbergGPT with benchmarked performance on key materials properties.
• The AGAPI-Agents framework orchestrates multistep scientific workflows through standardized API endpoints and asynchronous execution, ensuring reproducibility and scalable research.

The AtomGPT.org ecosystem is an open, modular platform for transformer-driven scientific discovery in materials science and quantum many-body systems. It unifies foundation models for atomistic property prediction, generative inverse design, and quantum system modeling, while integrating extensive toolchains, open-access APIs, high-throughput datasets, and agentic AI orchestration. AtomGPT.org is distinctive in its reliance on physics-aware data representations, open-source LLMs, and standard interfaces enabling rapid interoperability with simulation engines, databases, and workflow controllers. Core subdomains include the AtomGPT and RydbergGPT model families for materials and quantum systems, and the AGAPI-Agents orchestration layer for agentic execution across diverse scientific computing endpoints (Choudhary, 2024, Fitzek et al., 2024, Lee et al., 12 Dec 2025).

1. Foundational Model Architectures

AtomGPT.org’s model portfolio centers on transformer-based architectures, optimized for both forward property prediction and inverse structure generation. The canonical forward AtomGPT model is a fine-tuned GPT-2-small network (124 M parameters, 12 layers, 12 attention heads, hidden dimension 768), structurally identical to Vaswani et al. (2017) with standard sinusoidal or learned positional encodings. For property regression, the softmax head is replaced with a two-layer MLP (Linear(768→1024) → GELU → Linear(1024→1)), yielding scalar predictions for formation energies, electronic bandgaps, and superconducting critical temperatures (Choudhary, 2024).

Input representations begin with deterministic text encodings produced by ChemNLP or Robocrystallographer, encompassing chemical formula, space group, lattice metrics, bonding geometries, and explicit atomic coordinates. Tokens are mapped to 768-dimensional embeddings and further processed by the model. For generative tasks, AtomGPT employs quantized Mistral 7B LLMs with LoRA adapters, supporting parameter-efficient fine-tuning for domain-specific instruction following.

Within quantum simulation, RydbergGPT inaugurates the “atomic” foundation model suite. Its architecture couples a graph convolutional network (GCN) frontend (ingesting geometric and Hamiltonian graph attributes) with a vanilla encoder-decoder transformer core: a 2-layer GCN ($d_{\rm graph}=64$), a single-layer encoder ($d_{\rm model}=32$, $h=8$, $d_{\rm ff}=128$), and a three-block transformer decoder. All blocks leverage multi-head scaled-dot-product attention, layer normalization, and residual connections. The lightweight parameterization (~66 K trainable weights) ensures feasibility for single-GPU training and rapid experimentation (Fitzek et al., 2024).

2. Agentic AI Platform and Orchestration (AGAPI-Agents)

The AGAPI-Agents framework operationalizes agentic AI by decoupling reasoning (LLMs) from execution (tool endpoints) through a four-stage pipeline: Agent, Planner, Executor, and Summarizer (Lee et al., 12 Dec 2025). User queries submitted via web, SDK, or voice interface are transformed into explicit multi-tool plans. The Planner module constructs valid API call sequences based on system constraints and available tools, formalized as

$$P^{*} =\mathrm{Planner}\bigl(Q; \,\mathcal{T}, H\bigr) = \arg\max_{P\in\Pi}\, \mathbb{E}_{\mathrm{LLM}\Bigl[\mathrm{Utility}(Q,P)\Bigr]}$$

where $\mathcal{T}$ denotes tool schemas and $\Pi$ admissible plan space.

Eight open-source LLMs (e.g., Llama-3.2-90B-Vision, GPT-OSS-20B, Qwen3-Next-80B) are accessible via Ollama/vLLM deployment, tensor parallel serving, and version pinning. Tools are registered as OpenAPI 3.1 endpoints, with tightly defined request/response schemas validated by FastAPI/Pydantic. The Executor dispatches REST calls asynchronously, logs all transactions for reproducibility, and supports batching, error retries, and hybrid cloud/on-prem deployment. The Summarizer module collates results into structured outputs (Lee et al., 12 Dec 2025).

3. Data Integration, Simulation Engines, and ML Model Unification

API endpoints span over twenty materials science and quantum simulation tasks. These include database queries (/jarvis_dft/query), graph neural network inference (/alignn/query), force-field optimization (ALIGNN-FF), tight-binding band-structure calculation (SlaKoNet), interface and defect generation, and X-ray diffraction analysis (/pxrd/query) (Lee et al., 12 Dec 2025). Table-based organization of select endpoints:

Endpoint	Input Summary	Output Summary
/jarvis_dft/query	Filtered composition/properties	List of JARVIS-DFT IDs, structures, computed properties
/alignn/query	Structure, property list	Formation energy, bandgaps, elastic constants, etc.
/alignn_ff/query	Structure, max steps, tolerance	Relaxed structure, convergence, final energy
/pxrd/query	Structure, radiation, two-theta	Peak positions, intensities, pattern URL
/generate_interface	POSCARs of material A and B	Interface structure POSCAR

All tool schemas are injected into LLM prompts following the ReAct paradigm, ensuring agentic plans conform to valid tool argumentation. Underlying translation layers map REST calls to local scientific codes, database connectors, or ML inference servers. Repeated queries benefit from result caching, while full logs and deterministic seeds enable strict reproducibility (Lee et al., 12 Dec 2025).

4. Workflow Construction and Application Patterns

Typical workflows—comprising up to ten serial steps—can execute complex materials design and analysis pipelines entirely through agentic planning. For instance, a semiconductor defect analysis may involve selection and retrieval of a substrate structure, supercell generation, atom substitution, force-field relaxation, property prediction, bandstructure calculation, and report assembly. The dataflow can be formalized as

$$Q \xrightarrow{\text{Agent}} P=\{t_{1},...,t_{n}\} \xrightarrow{\text{Executor}} R=\{r_{1},...,r_{n}\} \xrightarrow{\text{Summarizer}} \widehat{R}$$

as detailed in AGAPI-Agents (Lee et al., 12 Dec 2025). Other cases include inverse design via AtomGPT (structure generation from property constraints), heterostructure synthesis, and powder XRD simulation. Execution time benchmarks show mean response times of 16.641 s under 1,000 concurrent users, with anticipated scaling improvements via container orchestration and model sharding.

5. Model Training, Benchmarking, and Validation

AtomGPT models are trained using JARVIS-DFT datasets (57,713 structures for formation energies and bandgaps, 1,058 for superconducting $T_c$, 18,167 for TB-mBJ bandgaps), with 80:10:10 train/val/test splits. Loss functions are mean absolute error (L₁ loss) for forward models and cross-entropy for generative/inverse models. Training is performed with Adam (forward) or AdamW (inverse), modest batch sizes (10), and up to 200 epochs on V100-class hardware (Choudhary, 2024).

Performance benchmarks reveal that AtomGPT is competitive with state-of-the-art graph neural networks (ALIGNN, CGCNN) on OptB88vdW bandgaps (MAE 0.139 eV), $T_c$ for superconductors (MAE 1.54 K), and formation energies (MAE 0.072 eV/atom), despite operating purely on text-derived representations. Inverse design outputs are validated through downstream high-throughput DFT (VASP/OptB88vdW) and electron–phonon coupling calculations (Quantum ESPRESSO) (Choudhary, 2024).

RydbergGPT benchmarks include learning projective measurement outcome distributions for Rydberg arrays, energy and order parameter inference, and generalization to unseen Hamiltonian parameters or system sizes. Groundstate predictions closely match QMC references near phase transitions, with limitations observed for high-temperature, mixed-state regimes due to the model’s positive-wavefunction ansatz (Fitzek et al., 2024).

6. Ecosystem Modularity, Interoperability, and Extension

AtomGPT.org enforces input/output standardization across all models and tooling. Data are exchanged as JSON/YAML (experiment descriptors, measurement outputs), with shared class interfaces for datasets, model encoding/decoding, and observable estimation. APIs abstract away the underlying model (e.g., model.encode(x) returns a $d_{\rm model}=32$0 tensor), facilitating swapping of Hamiltonian encoders, probability decoders, or token vocabularies to target different quantum systems (e.g., dipolar, ionic, superconducting qubits, Bose–Hubbard, spin chains) (Fitzek et al., 2024).

The platform is fully interoperable with HuggingFace LLMs and enables model replacement (LLaMA 2, Falcon, Mistral) through simple configuration changes. Underlying scientific tooling is built atop JARVIS-Tools, pymatgen, and connects to Materials Project, OQMD, NOMAD, AlphaFold, and other external scientific repositories (Choudhary, 2024, Lee et al., 12 Dec 2025).

AGAPI-Agents’ agentic control interfaces support extensibility via tool registry updates, new endpoint deployment, and modular summarization strategies. Access control, version pinning, reproducibility logging, and results leaderboard integration further promote transparency and research benchmarking.

7. Impact, Reproducibility, and Outlook

AtomGPT.org has seen adoption by over 1,000 researchers for end-to-end scientific workflows, supporting reproducible results via versioned models, deterministic sampling, and complete execution logs. Responses are rate-limited, authenticated (JWT), and validated for consistency and security. Case study benchmarks exhibit improved bulk modulus MAE (–27%) with tool augmentation, but declining predictive power for certain properties (e.g., $d_{\rm model}=32$1, bandgap) in tool-only settings, underscoring property dependence of workflow selection (Lee et al., 12 Dec 2025).

The AtomGPT.org approach—combining transformer-based LLMs with curated databases, GNN property predictors, physics engines, and agentic orchestration—lowers the barrier to AI-accelerated materials research, enables generative inverse design, and provides a scalable foundation for future expansion into new model families (e.g., IonGPT, HubbardGPT, SpinGPT), multi-property optimization, defect/interface analysis, and active learning pipelines (Choudhary, 2024, Fitzek et al., 2024, Lee et al., 12 Dec 2025).