DiffractGPT: PXRD Inverse Modeling

• DiffractGPT is a PXRD-based inverse model that converts powder X-ray diffraction data into detailed crystal structure information using Bragg’s law.
• It leverages a transformer backbone (Mistral-7B-Instruct with QLoRA) to generate precise lattice parameters, achieving a MAE of approximately 0.17–0.32 Å.
• Integrated within AGAPI, DiffractGPT enables automated diffraction analysis and tool-augmented workflows for inverse crystallography.

DiffractGPT is a diffraction-focused member of the AtomGPT family of modality-to-structure models for crystalline materials. In the formulation described alongside RamanGPT, DiffractGPT retains the same workflow and LLM backbone as the Raman inverse model, but replaces the Raman-spectrum prompt with a powder X-ray diffraction (PXRD) pattern; the physical link to structure is correspondingly more direct through Bragg’s law, and the reported lattice-parameter accuracy is higher, with lattice mean absolute error of approximately $0.17$–$0.32$ Å (Abel et al., 2 Jun 2026).

1. Definition and provenance

DiffractGPT is presented not as an isolated concept, but as a diffraction-specific instantiation of a broader bidirectional modality-structure framework. The relevant description appears in the discussion of RamanGPT, where the core approach is said to generalize to “any measurement with an unambiguous structural link,” and DiffractGPT is named as the PXRD case of that generalization (Abel et al., 2 Jun 2026).

Within the broader software ecosystem, DiffractGPT is also listed as one of the machine-learning tools integrated into AGAPI, an open-access agentic AI platform that connects open-source LLMs with materials-science APIs for workflows spanning materials data retrieval, graph neural network property prediction, force-field optimization, diffraction analysis, and inverse design (Lee et al., 12 Dec 2025). This places DiffractGPT at the intersection of inverse crystallography, diffraction informatics, and agentic scientific computation.

2. Inverse problem addressed by DiffractGPT

The central task associated with DiffractGPT is the inverse mapping from PXRD data to crystal structure. The governing premise, shared with related AtomGPT variants, is that serialized measurement data can be used to condition an instruction-tuned transformer that autoregressively generates crystallographic descriptions. In the DiffractGPT case, the measurement modality is PXRD rather than Raman spectroscopy, while the “workflow and LLM backbone are identical” to the Raman inverse model (Abel et al., 2 Jun 2026).

The key distinction is the physical relation between signal and structure. RamanGPT explicitly states that, for PXRD, “the mapping to structure is more direct (via Bragg’s law), yielding higher accuracy” (Abel et al., 2 Jun 2026). In encyclopedic terms, this means that DiffractGPT addresses an inverse problem that is still nontrivial, but is better constrained by crystallographic geometry than vibrational-spectrum inversion. The claim is methodological rather than rhetorical: the increased directness of the forward physics is tied to improved lattice reconstruction accuracy.

The same source also emphasizes the modularity of prompt design. It states that the approach generalizes by “serializing spectral/diffraction signals and conditioning instruction-tuned LLMs,” and that prompt conditioning can incorporate chemical formula, element list, or experimental context (Abel et al., 2 Jun 2026). This suggests a PXRD-conditioned generative workflow in which diffraction data are treated as structured sequence input rather than as an image-only or retrieval-only signal.

3. Architectural basis and generative workflow

Because DiffractGPT is described as having an identical workflow and LLM backbone to the Raman inverse model, its architecture is anchored in the same transformer-centered inverse-design recipe. The explicitly described backbone in RamanGPT is Mistral-7B-Instruct, fine-tuned with Quantized Low-Rank Adaptation (QLoRA) as a decoder-only transformer using scaled dot-product attention and autoregressive structure generation (Abel et al., 2 Jun 2026).

A direct implication is that DiffractGPT inherits the same general sequence-to-structure paradigm, with the measurement token stream changed from a binned Raman spectrum to a PXRD pattern. In the Raman inverse formulation, the model target comprises lattice constants, lattice angles, element list, and fractional atomic coordinates. The data block does not separately restate those targets for DiffractGPT, but the assertion of an identical workflow strongly indicates the same crystallographic output format (Abel et al., 2 Jun 2026).

The broader AtomGPT methodology further frames DiffractGPT as part of a bidirectional recipe in which graph neural networks handle structure-to-property prediction and LLMs handle property-to-structure generation. The RamanGPT summary states that “GNNs trained on property datasets (ALIGNN for spectra/DOS, etc.) and LLMs fine-tuned with sequence data comprise a general recipe for bidirectional modality-structure modeling,” with DiffractGPT given as the PXRD example of that generalization (Abel et al., 2 Jun 2026). This suggests that DiffractGPT is best understood not merely as a decoder from diffraction peaks to text, but as a diffraction-specialized inverse model embedded in a broader model-based materials pipeline.

4. Reported accuracy and the role of diffraction physics

The most concrete quantitative statement about DiffractGPT concerns lattice-parameter recovery. RamanGPT reports that, for PXRD, the mapping to structure is more direct via Bragg’s law and yields “higher accuracy (lattice MAE $\sim 0.17$–$0.32$ Å)” (Abel et al., 2 Jun 2026). The significance of this claim becomes clearer when placed beside the Raman inverse benchmark reported in the same work.

Modality instance	Structural relation stated in source	Reported lattice MAE
RamanGPT inverse model	Inverse mapping from Raman spectrum to structure	$1.14$–$2.16$ Å
DiffractGPT	Mapping to structure more direct via Bragg’s law	$\sim 0.17$–$0.32$ Å

This comparison indicates that DiffractGPT occupies the more strongly constrained end of the AtomGPT modality spectrum. The result does not imply that PXRD inversion is trivial; rather, it indicates that when diffraction geometry supplies a more explicit structure signal, an LLM-based inverse model can recover lattice parameters with markedly lower error than in the Raman case (Abel et al., 2 Jun 2026).

A common simplification is to treat all structure-from-measurement problems as equivalent sequence-learning tasks. The reported numbers argue against that simplification. In this framework, accuracy is not just a property of model scale or prompt engineering; it is strongly modulated by how directly the measurement encodes crystallographic periodicity.

5. Position within AGAPI and agentic diffraction workflows

AGAPI integrates DiffractGPT into a larger agentic infrastructure built around an Agent-Planner-Executor-Summarizer architecture. In that system, the LLM interprets the scientific query, the planner selects tools and execution order, the executor dispatches API calls, and the summarizer aggregates and formats the result (Lee et al., 12 Dec 2025). DiffractGPT appears in AGAPI’s list of integrated machine-learning models alongside ALIGNN, ALIGNN-FF, CHIPS-FF, and SlaKoNet.

AGAPI’s diffraction-analysis workflow is especially relevant to the operational role of DiffractGPT. The platform supports autonomous powder XRD analysis from natural-language query through structure retrieval, XRD computation, and result summarization. The workflow includes query parsing, structure retrieval from materials databases, XRD simulation setup, powder XRD computation, and output delivery through web, chatbot, or programmatic API interfaces (Lee et al., 12 Dec 2025). In this setting, DiffractGPT functions as part of a tool-grounded diffraction stack rather than as a free-running conversational model.

AGAPI also makes an important epistemic distinction: for diffraction analysis, tool access is “mandatory,” and LLMs “cannot synthesize full XRD patterns based solely on parametric data” (Lee et al., 12 Dec 2025). This directly addresses a recurrent misconception. DiffractGPT belongs to a tool-augmented scientific workflow in which structural reasoning is coupled to explicit diffraction computation and database retrieval, not to unconstrained language-model extrapolation.

6. Extensions, scope, and interpretive limits

DiffractGPT is presented as one instance of a modality-general approach rather than as the terminus of that approach. RamanGPT explicitly notes potential extensions to infrared, terahertz, inelastic neutron spectra, or multi-modal combinations such as Raman+PXRD, achieved by altering the database and prompt template while retaining the core strategy of conditioning instruction-tuned LLMs on serialized measurements (Abel et al., 2 Jun 2026). A plausible implication is that DiffractGPT can serve as the diffraction anchor in multimodal inverse crystallography, where PXRD constrains lattice geometry and complementary probes resolve ambiguities that diffraction alone may leave open.

The same evidence base also implies clear limits. Performance is coupled to domain coverage, prompt design, and the directness of the modality-structure relationship. The superior lattice accuracy reported for DiffractGPT arises precisely because PXRD is more directly linked to crystal structure through Bragg’s law; it should not be generalized to all spectroscopy- or scattering-based inverse problems without qualification (Abel et al., 2 Jun 2026).

In the current open-access materials-AI ecosystem, DiffractGPT is therefore best characterized as