AI in Chemical Calculations: Publishing Faster Without Losing Credibility
MatterGen, machine learning force fields, and the specific risks that researchers need to manage before submission.
In January 2025, Microsoft Research published a paper in Nature introducing MatterGen — and the materials science community has been processing the implications ever since. The team introduced MatterGen as a diffusion-based generative model designed to create stable and diverse inorganic materials across the periodic table. Unlike traditional image-based diffusion models that rely on adding Gaussian noise, MatterGen employs a customised diffusion process tailored to the periodic structure and symmetries of crystalline materials, with each component undergoing a corruption process that respects the material's geometry.
Published in Nature, the study demonstrates MatterGen's ability to explore uncharted material spaces, outperform traditional methods in identifying stable, unique, and novel materials, and deliver experimentally validated results with real-world applications in energy, carbon capture, and semiconductors.
This is, by any reasonable assessment, a genuinely significant development. But MatterGen is also a useful illustration of the broader challenge this article is concerned with: as AI tools become capable of accelerating chemical calculations by orders of magnitude, the question of how to publish that work credibly, transparently, and in ways that will survive peer review and downstream replication has become one of the most practically important questions in computational chemistry.
The tools are extraordinary. The risks of using them carelessly are real and specific. And the gap between what researchers are being told about using AI in their science and what journals are actually requiring of them is wider than it should be.
What MatterGen and Machine Learning Force Fields Actually Do
Before examining the risks, it is worth being precise about what these tools offer — because the benefits are genuine and the pace of their adoption in serious research is accelerating rapidly. The AI-in-chemistry landscape has two distinct poles, with different functions and different uncertainty profiles.
Designs novel materials from scratch rather than screening known candidates. Given prompts for chemistry, mechanical, electronic, or magnetic properties — or combinations of constraints — it generates materials beyond the limited set of known ones, enabling efficient exploration of uncharted chemical space.
Rather than generating new structures, machine learning force fields predict the energies and forces governing the behaviour of known or hypothetical structures — enabling molecular dynamics simulations at a fraction of the computational cost of density functional theory (DFT).
MLFFs trained on quantum-mechanical energies and forces have shown the capacity to achieve sub-kcal mol⁻¹ Å⁻¹ accuracy while maintaining computational efficiency. In practical terms, this means that simulations which previously required supercomputer time measured in weeks can now be run on a standard workstation in hours. For a research group at a university in Georgia or Kazakhstan with limited access to high-performance computing infrastructure, this is not a minor convenience — it is a fundamental change in what science is possible.
The result has been a dramatic acceleration in the rate at which computational chemistry papers can be produced. The question is whether that acceleration is producing reliable science at commensurate pace.
Where the Risks Live
A significant risk occurs when MLFFs do not fail outright but produce seemingly reasonable yet ultimately incorrect results. Identifying such cases requires the kind of system-specific expertise that many researchers using these tools — particularly experimentalists adopting computational methods — may not have developed. The risks fall into three distinct categories.
-
Plausible-but-wrong results outside the training domain
The TEA Challenge 2023, a systematic evaluation of MLFF performance, found that the main discrepancies between molecular dynamics results occur when the training datasets lack representative reference geometries. In plain terms: if the chemistry you are simulating is similar to the training data, you are likely to get reliable results. If you are working in a region of chemical space the model has not seen, you may get results that look plausible and are not. The model does not tell you which situation you are in — it simply produces numbers.
-
Benchmark accuracy that does not transfer to your system
Disconnects between simulation stability and mechanical property accuracy have been observed, with prediction errors correlating with training data representation rather than the modelling method. Current computational benchmarks provide valuable controlled comparisons, but they may overestimate model reliability when extrapolated to experimentally complex chemical spaces. A paper reporting MLFF results without demonstrating that the model performs reliably in the specific chemical regime studied is making an implicit claim that may not hold — and that a careful reviewer will be right to challenge.
-
Reactivity and transition states in catalysis
Reliable transferability in chemical space, reactivity, and scalability in terms of system size have yet to be fully addressed. For catalysis specifically — a field where reaction mechanisms involve bond-breaking and bond-forming events in poorly represented chemical environments — this limitation is particularly significant. An MLFF that performs well for bulk property prediction of stable materials may produce qualitatively incorrect results for a transition state or an adsorption event at a catalyst surface.
The benchmark accuracy of a model on standard test sets does not reliably predict its accuracy on your specific system.
The Disclosure Landscape in Computational Chemistry
The question of what researchers are required to disclose when using AI tools in chemical calculations is genuinely unsettled, and the gap between practice and requirement is wider in computational chemistry than in almost any other field where AI is now being deployed.
While many journals have introduced mandatory disclosure sections regarding AI use in manuscript preparation, similar measures have yet to be implemented consistently for results generation. For computational chemistry, the disclosure question is more complex than for writing assistance — because the AI tool is not helping communicate the science, it is generating the scientific results themselves.
The distinction matters enormously for credibility. Using ChatGPT to improve the clarity of a methods section is a question of writing assistance. Using MatterGen to generate the candidate structures that form the basis of an experimental campaign, or using an MLFF to produce the molecular dynamics data that underlies a paper's core claims, is a methodological choice as significant as the choice of any other computational method — and needs to be documented with the same rigour.
Current best practice, as it is emerging across journals including those in the Nature Portfolio and the ACS family, requires that AI tools used in the scientific workflow be disclosed in the Methods section with sufficient specificity to allow reproduction: the tool name and version, the training dataset used if known, the hyperparameters of any fine-tuning, and the validation approach used to confirm that the model is performing reliably in the relevant chemical regime.
Authors must include explicit acknowledgement of model limitations, uncertainties, and potential biases, including those inherent in training data, as well as mitigation strategies for identified biases. All claims about model capabilities must be supported by empirical evidence from the study; unsubstantiated assertions must be avoided.
This standard is demanding — deliberately so. A paper that uses a machine learning force field without demonstrating its accuracy in the specific chemical system studied is making a claim that its results are physically meaningful without providing the evidence necessary for a reviewer to assess that claim. This is precisely the kind of gap that produces retraction notices six months after publication, when an independent group discovers that the results cannot be reproduced because the model was operating outside its valid domain.
What Credible AI-Assisted Chemistry Papers Look Like
The papers that successfully integrate AI tools into computational chemistry while maintaining credibility — and surviving peer review at serious journals — share a set of common features worth examining as a practical template.
Five features of a credible AI-assisted chemistry paper
-
They benchmark the AI tool against DFT for the specific system
Rather than relying on published benchmark performance, credible MLFF papers compare predictions to DFT calculations for a representative subset of the configurations studied. This does not require DFT for every configuration — which would defeat the purpose — but it provides evidence that the model performs reliably in the relevant regime.
-
They are explicit about the training data domain
A paper using an MLFF trained on oxide surfaces should explicitly address whether and how the model was validated for the specific oxide composition and surface termination studied. Universal models like MACE-MP-0 and CHGNet have known limitations in specific chemical environments that need to be acknowledged rather than assumed away.
-
They separate generative and predictive AI contributions clearly
In a workflow that uses MatterGen to generate candidate structures and an MLFF to predict their properties, these are two distinct AI contributions with different uncertainty profiles. A credible paper treats them as such, rather than presenting the combined output as a single validated calculation.
-
They validate predictions with experiment where possible
MatterGen's Nature paper explicitly includes experimentally validated results — the gold standard, and not coincidentally appearing in a top-tier journal. Where full experimental validation is not possible, partial validation of key predictions and explicit acknowledgement of what remains unvalidated are the minimum credible standard.
-
They use uncertainty quantification
Several modern MLFF architectures — including MACE and NequIP — can provide uncertainty estimates alongside predictions, flagging configurations where the model is extrapolating beyond its training domain. Using these estimates and reporting them is both scientifically sound and increasingly expected by reviewers familiar with the tools' limitations.
The Practical Upside: What These Tools Genuinely Enable
It is worth being direct about what rigorous use of these tools actually makes possible — because the answer for researchers at smaller institutions with limited computational resources is significant.
A research group in Almaty or Tbilisi that ten years ago could not access the DFT calculation infrastructure to study catalytic mechanisms at realistic timescales can now, using publicly available MLFF models, run microsecond molecular dynamics simulations on a standard workstation. A group studying battery electrode materials that previously required months of supercomputer allocation to screen candidate compositions can now use MatterGen to generate novel candidates and an MLFF to pre-screen their stability properties, reducing the experimental burden to the most promising hits.
In the beginning, maybe one or two out of ten candidates will be successful, but this will gradually improve as people build more trust in these generative models — precisely what has been happening in generative AI for small molecule drug discovery over the past few years. The democratisation of computational chemistry capability is real and it is consequential for researchers who have historically been disadvantaged by their distance from major computing facilities.
The credibility framework described above is not a barrier to using these tools. It is the condition under which using them produces science that the community can trust, build on, and fund. A well-validated AI-assisted computational chemistry paper submitted to ACS Catalysis or the Journal of Physical Chemistry Letters is both faster to produce and more credible than a poorly documented one — because the validation work it contains is itself a scientific contribution, not merely a formality.
The Bottom Line
The tools are genuinely powerful. The validation requirements are clear. The gap between them, for researchers who want to publish faster without publishing things that will not survive scrutiny, is exactly the gap that rigorous methodology closes.
The acceleration that MatterGen and machine learning force fields offer is real, and for researchers far from major computing centres it is genuinely levelling. But acceleration without validation is not speed — it is risk, deferred to the moment six months after publication when someone tries to reproduce the result. The papers that will still be standing then are the ones that did the validation work before submission, not after the retraction notice.
Questions readers ask after this piece
What is MatterGen and how is it different from traditional materials screening?
MatterGen is a diffusion-based generative model from Microsoft Research, published in Nature in January 2025, designed to create stable and diverse inorganic materials across the periodic table. Instead of screening known candidates, it directly generates novel materials given prompts of the design requirements — chemistry, mechanical, electronic, or magnetic properties, or combinations of constraints — enabling exploration of materials beyond the limited set of known ones.
What is a machine learning force field (MLFF) and why is it useful?
A machine learning force field predicts the energies and forces governing the behaviour of structures, enabling molecular dynamics simulations at a fraction of the computational cost of density functional theory (DFT). MLFFs trained on quantum-mechanical data can achieve sub-kcal/mol/Å accuracy while remaining efficient — simulations that previously required weeks of supercomputer time can run on a standard workstation in hours, which is transformative for resource-limited research groups.
Why does benchmark accuracy not guarantee reliable results for my specific system?
The benchmark accuracy of an MLFF on standard test sets does not reliably predict its accuracy on a specific system. The TEA Challenge 2023 found that the main discrepancies in molecular dynamics results occur when training datasets lack representative reference geometries, and that prediction errors correlate with training data representation rather than the modelling method. If you are working in a region of chemical space the model has not seen, it may produce plausible but incorrect numbers — and it will not tell you which situation you are in.
What do I need to disclose when using AI tools to generate chemical results?
Emerging best practice across Nature Portfolio and ACS journals requires disclosure in the Methods section with enough specificity to allow reproduction: the tool name and version, the training dataset if known, the hyperparameters of any fine-tuning, and the validation approach used to confirm the model performs reliably in the relevant chemical regime. Authors must also explicitly acknowledge model limitations, uncertainties, and biases, and support all capability claims with empirical evidence from the study.
What does a credible AI-assisted computational chemistry paper look like?
Credible papers share five features: they benchmark the AI tool against DFT for the specific system studied; they are explicit about the training data domain and its limitations; they separate generative and predictive AI contributions clearly; they validate predictions against experiment where possible; and they use and report uncertainty quantification to flag configurations where the model is extrapolating beyond its training domain.
Are MLFFs reliable for catalysis and reaction mechanisms?
With particular caution. Reliable transferability in chemical space, reactivity, and scalability remain incompletely addressed. For catalysis — where mechanisms involve bond-breaking and bond-forming in poorly represented chemical environments — an MLFF that performs well for bulk property prediction of stable materials may produce qualitatively incorrect results for a transition state or an adsorption event at a catalyst surface. System-specific validation against DFT is essential before publishing such results.