Six Open Questions in Machine-Learned Interatomic Potential Foundation Models

This paper defines foundational machine-learned interatomic potentials (MLIPs) and articulates six critical open questions that are expected to guide future cutting-edge research in the field.

The Gist

Imagine you are trying to predict how a crowd of people will move, bump into each other, and react to a sudden push. In the world of atoms, scientists use "Interatomic Potentials" to do exactly this: they calculate how atoms push and pull on one another to predict how materials behave.

For decades, scientists had to build a custom "rulebook" for every single type of material (like a rulebook just for gold, another just for water, another just for steel). These rulebooks were accurate but took years to write and couldn't be used for anything else.

Recently, a new type of AI called Machine-Learned Interatomic Potentials (MLIPs) has arrived. Even better, we now have "Foundation Models." Think of these as a "Super-Grandmaster" AI that has read every chemistry textbook in the library. It hasn't just memorized one rulebook; it has learned the general language of matter. Now, if you ask it about a new material it has never seen before, it can guess the rules with very little extra training.

However, the authors of this paper argue that while this technology is exciting, we are asking the wrong questions or not asking the right ones yet. They have identified six big open questions that scientists need to solve before these AI models can truly revolutionize science.

Here are the six questions, explained with simple analogies:

1. What actually counts as a "Foundation Model" for atoms?

The Analogy: Imagine a chef who can cook a perfect steak. That's a specialist. Now imagine a chef who can cook a steak, bake a cake, brew coffee, and grill a fish, all without needing a new recipe book for each one. That's a "foundation model."
The Question: We need to agree on the minimum requirements. Does the AI just need to be good at many things? Or does it need to be able to learn new tasks instantly? The paper suggests we need a clear definition so we don't just call any good AI a "foundation model" when it's actually just a narrow specialist in disguise.

2. Do we need more data, better data, or smarter models?

The Analogy: Imagine trying to teach a child to recognize dogs.

• More Data: Showing the child 1 million pictures of dogs.
• Better Data: Showing the child 1,000 perfect pictures of dogs from every angle, in every weather, with no blurry photos.
• Smarter Models: Giving the child a better brain (or a better way of thinking) so they can learn from fewer pictures.
  The Question: The paper asks: Should we just dump more data into the AI? Or should we spend time curating "perfect" data? Or should we build smarter AI brains that can learn from less? The answer isn't simple; it's likely a mix of all three, but we don't know the perfect recipe yet.

3. Can these AIs handle "long-distance" relationships?

The Analogy: Imagine a crowded room. If you push someone, the person right next to you feels it immediately. But what about the person across the room? In physics, atoms can "feel" each other across distances (like magnets or static electricity).
Most current AI models are like people who only talk to their immediate neighbors. They are great at local gossip but terrible at understanding the vibe of the whole room.
The Question: Can these models learn to "hear" the whispers from across the room? The paper notes that for some materials (like charged crystals), ignoring the long-distance whispers leads to wrong answers. We need to know if the AI can fix this without becoming too slow to use.

4. Can the AI discover new physics, or is it just guessing?

The Analogy: Imagine a student who has studied every past exam. If you give them a new question that looks exactly like an old one, they will ace it. But if you ask a question about a concept that was never in the book, will they make a logical guess, or will they just hallucinate a fake answer?
The Question: Can these AIs look at a strange, high-pressure situation (like the center of a planet) and say, "I've never seen this, but based on the laws of physics I've learned, I think this will happen"? Or are they just memorizing patterns? The paper is skeptical; currently, they are mostly very good at interpolation (filling in the blanks) but bad at true discovery.

5. Can they scale up to do useful simulations?

The Analogy: A super-fast sports car is great for a short track. But if you want to drive a cross-country truck, you need something that can carry a heavy load without running out of gas.
The Question: The most accurate AI models are often so heavy and slow that they can only simulate a tiny speck of dust for a tiny fraction of a second. The paper asks: Can we make these models fast enough to simulate a whole virus, a battery, or a piece of metal for a long time? If the AI takes longer to run than the supercomputer it's running on, it's not useful.

6. How do we know if the AI is actually good?

The Analogy: Imagine a video game leaderboard. If everyone just plays the same level over and over to get the highest score, the leaderboard stops telling you who is actually the best player. They might just be "cheating" the specific test.
The Question: We have a popular "test" (called Matbench Discovery) that ranks these AI models. But the paper warns that if everyone trains their AI specifically to pass that one test, the scores will get stuck at the top, and we won't know if the models are actually improving in real life. We need better, more diverse tests that catch the AI when it tries to cheat or when it fails in real-world scenarios.

The Bottom Line

The paper concludes that we are in a "Gold Rush" moment for this technology. We have powerful new tools (Foundation Models) that promise to let us design new medicines, batteries, and materials from scratch. But before we get too excited, we need to stop and ask: Are these tools actually ready?

The authors aren't saying the technology is bad; they are saying it's too new and fast-moving. We need to define what it is, fix its blind spots (like long-distance interactions), make it faster, and create better tests to ensure it's not just memorizing answers but actually learning the laws of nature.

Technical Summary

Technical Summary: Six Open Questions for MLIPs

Problem Statement

Machine-learned interatomic potentials (MLIPs) have rapidly evolved from proof-of-concept research to a central tool in atomistic modelling, promising to resolve the historical trade-off between simulation scale and accuracy. Recently, the paradigm has shifted toward "foundation models"—large, pre-trained models trained on diverse datasets capable of simulating multiple chemistries with minimal adaptation. However, the field is characterized by a proliferation of new architectures and designs, creating a fragmented landscape where specialists struggle to maintain a coherent overview.

The core problem addressed by this paper is the lack of consensus on the fundamental definitions, capabilities, and limitations of these emerging "foundation" MLIPs. While progress is rapid, critical questions regarding their theoretical underpinnings, data requirements, physical fidelity (particularly regarding long-range interactions), and scalability remain unresolved. The authors argue that without articulating these open questions, the field risks developing incoherent standards and misinterpreting the true capabilities of these models.

Methodology

This paper does not present new experimental data, novel algorithms, or a specific benchmark study. Instead, it employs a conceptual and analytical framework derived from the "ML4Atoms" reading group and input from leading MLIP developers. The methodology involves:

1. Taxonomic Adaptation: The authors adapt the definition of "foundation models" from the machine learning community (specifically the framework by Bommasani et al.) to the domain of atomistic modelling. They define five criteria: expressivity, scalability, memorization, multimodality, and compositionality.
2. Literature Synthesis: The paper synthesizes current state-of-the-art research, architectural trends (e.g., equivariant GNNs, transformers, ACE), and empirical scaling laws to frame six specific open questions.
3. Critical Analysis: Each of the six questions is explored by reviewing existing literature, identifying contradictions in the field (e.g., conflicting reports on long-range interaction handling), and highlighting gaps between current capabilities and the theoretical requirements of a true foundation model.
4. Visualization: The authors utilize vectorized text analysis and cosine similarity to map the relationships between the definition criteria and the open questions, illustrating how these concepts interconnect.

Key Contributions

The paper's primary contribution is the articulation and structuring of six critical open questions that define the current frontier of MLIP research:

1. What is the minimal definition of an atomistic foundation model?
   The authors propose a working definition based on five adapted criteria:

   • Expressivity: The capacity to capture complex physical interactions (e.g., body-order, equivariance).
   • Scalability: The ability to improve performance with increased data, parameters, or compute, adhering to scaling laws.
   • Memorization: The ability to retain learned capabilities while adapting to new domains via fine-tuning.
   • Multimodality: Reinterpreted as multi-fidelity learning (integrating data from different levels of theory, e.g., PBE vs. hybrid functionals).
   • Compositionality: The ability to accurately describe large systems composed of familiar local units (e.g., polymers from monomers).

2. Do we need more data, better data, or better models?
   The paper argues that progress is not linear in any single dimension. It highlights that "brute-force" scaling of data volume is often inefficient due to noise in high-throughput datasets (e.g., mixed PBE/PBE+U). Instead, the field requires a coordinated advance in model complexity (inductive biases like equivariance), data diversity (covering non-equilibrium states), and data quality (high-fidelity targets for specific properties).

3. Can MLIPs really handle long-range interactions, and does it matter?
   The authors distinguish between physics-based definitions (e.g., Coulombic tails, correlation length) and graph-based definitions (influence propagation). They note that while local message-passing neural networks (MPNNs) suffer from over-smoothing and under-reaching, various strategies exist to mitigate this, including explicit physics terms (Ewald summation, charge equilibration) and architectural innovations (attention mechanisms, virtual nodes). A unified theoretical framework for long-range MLIPs remains elusive.

4. Can MLIPs discover truly new physics?
   The paper distinguishes between generalizability (extrapolating to new regimes) and discovery (uncovering new principles). Current MLIPs often struggle with rare events and extreme conditions due to training biases toward near-equilibrium states. While some "emergent" behaviors (e.g., stable liquid dynamics from crystal-trained models) have been observed, the field lacks robust metrics to distinguish genuine physical discovery from complex statistical interpolation.

5. Can MLIPs scale to do more useful simulations?
   Scalability is identified as a bottleneck. Large foundation models often incur high computational costs and memory usage, limiting their application to systems >50k atoms. The paper reviews strategies to improve efficiency, including knowledge distillation (compressing large models into smaller, specialized ones), hardware-specific optimizations (e.g., cuEquivariance, FlashAttention), and mixture-of-experts (MoE) architectures.

6. How do we know if our MLIPs are any good?
   The paper critiques current benchmarking practices, noting the dominance of the "Matbench Discovery" leaderboard. It warns against the "McNamara fallacy" (focusing only on easily measurable metrics) and "Goodhart's Law" (metrics becoming targets). The authors advocate for a spectrum of benchmarks ranging from representational (architecture-level properties) to pragmatic (task-specific utility) to ensure models are evaluated on their actual scientific utility.

Results and Findings

As a review and perspective paper, the authors do not present quantitative results of a new model. Instead, their "results" are the synthesis of the current state of the field:

• Definition Gap: There is no consensus on what constitutes a "foundation" MLIP; current models satisfy subsets of the proposed criteria but rarely all simultaneously.
• Data vs. Model Trade-off: Increasing dataset size alone does not guarantee better transferability; model expressivity and data quality (fidelity/diversity) are often the limiting factors.
• Long-Range Ambiguity: While methods exist to handle long-range interactions (explicit physics vs. learned attention), there is no single approach that universally outperforms others across all physical regimes.
• Extrapolation Limits: Current models are largely interpolative. While some emergent behaviors are observed, the ability to autonomously discover new physics or handle extreme conditions (high pressure/temperature) remains limited and often unreliable.
• Scalability Barriers: Computational cost remains a significant barrier for large-scale simulations, necessitating innovations in model compression and hardware utilization.
• Benchmark Saturation: Current leaderboards show signs of saturation, where incremental gains may reflect noise rather than genuine improvement, highlighting the need for more diverse and pragmatic evaluation metrics.

Significance

The paper positions itself as a critical "stock-taking" exercise for a rapidly maturing field. Its significance lies in:

• Providing a Conceptual Framework: By adapting the "foundation model" taxonomy to atomistic science, it offers a structured vocabulary for discussing MLIP capabilities and limitations.
• Guiding Future Research: The six questions serve as a roadmap for the community, identifying where research investment is most needed (e.g., unified long-range theories, better benchmarks, efficient scaling).
• Promoting Critical Engagement: The authors explicitly aim to avoid "easy acceptance of orthodoxy," encouraging the community to critically examine assumptions about data scaling, model architecture, and physical fidelity.
• Bridging Disciplines: It connects the fast-moving developments in ML (transformers, foundation models) with the rigorous physical constraints of materials science, ensuring that MLIP development remains grounded in physical reality.

The paper concludes that while foundation MLIPs hold the promise of enabling bottom-up atomistic design of new materials and chemistry, realizing this potential requires addressing these fundamental open questions through intense and open examination.