MACE and Scientific AI: What a Rotation Test Can Tell You

Turn a model of a molecule on your desk. You have changed the direction it faces, rather than the arrangement of its atoms relative to one another. A scientific AI system should distinguish that change of viewpoint from a change to the structure itself.

This is a useful entry point into MACE, a neural-network architecture for interatomic potentials. The subject can sound remote from everyday AI use. Yet it exposes a question that matters whenever we evaluate a model: which changes in an input should change its answer, and which should leave it alone?

Getting that relationship right is valuable. It is also only one part of getting the science right.

What MACE was designed to learn

Ilyes Batatia and colleagues presented MACE at NeurIPS 2022. Their work combines equivariant message passing with higher-order messages to model atomic systems. In this setting, a graph represents atoms and their local neighbourhoods, and the model predicts contributions to the system’s energy.

Higher-order messages are a way of representing information involving groups of atoms, rather than relying only on pairwise messages. The paper investigates how that choice affects accuracy, learning curves and computational cost on named molecular benchmarks. These are research results for the reported configurations and tests, not a claim that any model carrying the MACE name is suitable for every material.

The architecture is especially interesting as an example of putting structure into a learning system. The model is not simply asked to discover every relevant relationship from an unrestricted collection of examples.

An unchanged number and a rotated direction

The distinction between invariance and equivariance is explained in the 2022 NequIP research paper, an earlier contribution to this area. For the atomic systems considered, total potential energy is invariant under rotating the whole structure. It remains the same scalar quantity. Forces are directional: their vectors should rotate with the structure.

That coordinated change is equivariance. It does not mean every output stays identical. It means the output changes in the corresponding, specified way when the input is transformed. The NequIP paper also describes obtaining forces from the negative gradient of predicted energy.

For a non-specialist, the practical distinction is enough: do not ask whether a model is generally ‘unchanged by rotation’. Ask which output you are inspecting. A number describing energy and an arrow describing force have different obligations.

Changing orientation is not changing the experiment

Consider a hypothetical demonstration with one stored atomic configuration. A presenter calculates an energy, rotates every atomic coordinate together and calculates it again. They report that the two energy predictions agree within an appropriate numerical tolerance.

That is an intelligible consistency check. Now imagine they move one atom relative to the others and obtain a different energy. That second change is not a failure of the first principle. The input now describes a different configuration.

The same care applies to directional outputs. A force arrow that stays fixed on the screen while the entire structure rotates could be the wrong behaviour. An arrow that rotates appropriately could be the right behaviour, even though its displayed coordinate components have changed.

These examples are explanatory scenarios, not tests performed for this article. They show why a demonstration should state exactly what was transformed and what was expected to happen.

Consistency cannot supply the missing reference

Suppose a deliberately useless model returns the same energy number for every possible configuration. It would pass the simple rotation check in our hypothetical demonstration. That would give us no reason to trust its predictions when the atomic arrangement changes.

This thought experiment separates two questions. Does the output respect the required transformation? Does it agree closely enough with suitable reference evidence for the intended task? A convincing answer to one cannot replace an answer to the other.

For someone reading a research announcement, that means looking beyond an attractive animation of a rotating molecule. Ask what the model was trained on, what was held out for evaluation and which quantities were compared. Then ask how far the proposed use departs from those tests.

A similar distinction appears in MaryChuks.com’s discussion of confidence at a predicted protein interface: the evidence has to match the particular claim. Different scientific models call for different checks, even when their visualisations look persuasive.

Read the benchmark as a bounded claim

The original MACE paper reports experiments on molecular benchmark tasks, including rMD17 and 3BPA. It examines both predictive performance and the effect of architectural choices. Read those details as part of the result rather than as optional small print.

If a hypothetical materials business proposes using a related model for a new substance, a useful review would identify the exact model and training data being proposed. It would specify the configurations the business expects to encounter and the reference comparisons it plans to make.

The business should also name the decision the prediction will support. Is it organising candidates for further investigation, exploring a mechanism, or contributing to a claim about a product? Those uses imply different consequences when a prediction is wrong.

This is an editorial framework for reviewing an application, not an assessment that a particular deployment is ready. The 2022 paper alone cannot settle the suitability of a later, differently trained system.

Ask for a demonstration you can interpret

A useful explanation can present the transformation check beside the accuracy check, with a clear account of what each establishes. The presenter should identify units, numerical tolerances and the scope of the reference comparison.

When discussing speed, request the workload and computing setup behind the comparison. When discussing generalisation, ask what changed between training and testing. Avoid converting a qualified experimental result into an unqualified headline about understanding physics.

Independent creators covering scientific AI can use these questions without pretending to reproduce an entire research programme. Explain the tested relationship, preserve its limits and distinguish the researchers’ findings from the applications you are proposing.

MACE illustrates an important direction in AI design: choosing mathematical relationships that the system should preserve. The reader’s task is to recognise the value of that structure while still asking for evidence that the resulting predictions answer the scientific question at hand.

Sources

Batatia and colleagues, MACE, NeurIPS 2022; Batzner and colleagues, E(3)-equivariant graph neural networks for data-efficient and accurate interatomic potentials, Nature Communications, 4 May 2022. Checked 7 November 2026. Demonstrations and business scenarios above are illustrative editorial analysis.