On 5 July, an arXiv preprint showed that optimizers called SOAP and SOAP-Muon train the AI models behind molecular and materials simulation substantially faster and more accurately than Adam — the default the field has used for years [S1]. The gains are biggest exactly where researchers hurt most: when labelled data is thin. Why has everyone been reaching for the same tool, and what does that overlooked choice unlock?

The tool nobody tested

Machine learning interatomic potentials, or MLIPs, are neural networks that predict how atoms interact — replacing quantum-mechanical calculations that would take days with predictions that take milliseconds [S1]. The authors describe them as a defining feature of AI-driven scientific simulation [S1], underpinning work in catalysis, battery design, and drug discovery.

But while the architectures of these models have evolved rapidly, one design choice has stayed frozen: the optimizer — the algorithm that adjusts the model's weights during training. The community has defaulted to Adam and its variants, and the authors note that this decision has received little scrutiny [S1].

Adam is the workhorse optimizer of deep learning. It decides how big a step to take each time the model updates its parameters, and it has been the default across most of AI since around 2015. Reaching for it is understandable. Questioning it has been rare.

What it means

The authors — Gil Harari, Yoel Zimmermann, Ola Tangen Kulseng, Laura Zichi, Chuin Wei Tan, Marc L. Descoteaux, and Boris Kozinsky [P2] — implemented and systematically compared a class of matrix-structured optimizers for training two well-known MLIP architectures, NequIP and Allegro [S1]. These optimizers — Muon, SOAP, and a hybrid called SOAP-Muon — take a fundamentally different approach to weight updates. Instead of treating each parameter independently, as Adam does, they operate on whole matrices of parameters at once, capturing structural relationships that Adam misses.

The result: these optimizers substantially outperform Adam in both convergence speed (how quickly the model stops improving) and final accuracy (how good the model gets) [S1]. SOAP and the SOAP-Muon hybrid emerged as the robust, consistently strong performers across the experiments [S1]. Muon alone delivered only partial gains over Adam [S1] — a useful reminder that not every new optimizer is a winner.

The most striking finding is where the gains were largest: under partial force supervision [S1]. In MLIP training, models learn from both energy labels and force labels — forces being the gradients that tell you how atoms will move. Force labels are expensive to generate because they require quantum-mechanical calculations. Partial force supervision means training with only some of those labels available. When data is scarce, the matrix-structured optimizers shone brightest — which matters because labelled data is the single biggest bottleneck in MLIP development.

The authors conclude that the choice of optimizer is a design dimension the field has neglected, despite its significant effect on performance [S1] — a polite way of saying the field has been leaving performance on the table.

What it means for business

For the labs and companies building MLIPs — from materials-science startups to pharmaceutical teams running molecular simulations — the practical implication is straightforward: you may be able to train better models with less labelled data, or train comparable models faster, simply by swapping the optimizer.

A two-person computational chemistry team that currently spends weeks generating quantum-mechanical force labels to train an Allegro model could, if these results hold, reach the same accuracy with fewer labels by switching to SOAP or SOAP-Muon. That translates directly to compute cost — fewer density-functional-theory calculations, fewer GPU-hours generating training data.

For a contract simulation shop or a suburban materials-research firm, the value proposition is similar: faster convergence means more model iterations per quarter, more materials screened, more candidates evaluated.

The caveat is real. These results come from a preprint, not yet peer-reviewed [S1], and were tested on only two MLIP architectures [S1]. The optimizers are also more complex to implement than Adam, which ships with every major deep-learning framework. Adoption will depend on whether the community reproduces the results and whether the optimizers get integrated into standard MLIP libraries.

What we don't know yet

  • No specific numbers. The abstract reports substantial improvements but gives no percentage speed-ups, accuracy scores, or training-step counts [S1]. The full paper, available on arXiv [P2], likely contains these, but the headline claims remain unquantified in the abstract.
  • Generalisability is unproven. Only NequIP and Allegro were tested [S1]. Whether SOAP and SOAP-Muon deliver the same gains on other MLIP architectures — MACE, NequIP's successors, or custom in-house models — remains an open question.
  • Peer review is pending. The paper is listed on OpenReview as an ICML 2026 AI4Science Oral [P4], a strong signal of quality, but it is still formally a preprint [S1]. The review process may surface limitations the authors did not address.
  • The claim that the community defaults to Adam is author framing, not an independently verified survey [S1]. It is almost certainly true — Adam is the default in nearly every MLIP codebase — but should be read as informed observation, not measured fact.

The next concrete event to watch: the ICML 2026 AI4Science session, where this work is scheduled for an oral presentation [P4]. That is where reviewers' questions will surface and where the community gets its first chance to probe the results in public.

If this piece sharpened your thinking, subscribe — we'll be watching what comes out of that session.

Sources

Generated from an audited evidence pack with primary-source research. Social-media items are discussion signals, not verified facts. Nothing here is financial, legal or medical advice.