AI for molecular simulations may not need built-in physics to deliver strong results

Phys · Apr 22, 2026 · 555 words · By Jean-Paul Olivier

headphones Listen to the eFinder podcast briefing

Generate a natural audio summary of this story

Daily briefing

What to know about AI for molecular simulations may not need built-in physics to deliver strong results

The article discusses a novel machine learning approach, MD-ET, that simulates molecular dynamics without explicitly encoding fundamental physical laws like energy conservation or equivariance. The model, trained on a large dataset, achieved state-of-the-art results on some benchmarks, suggesting that physical behavior might be learnable from data alone. However, the authors caution that the unconstrained approach shows limitations, particularly regarding energy conservation in NVE simulations.

Propaganda risk 0%

Claims checked 12

Techniques found 0

Topics 0

Coverage spectrum

Coverage gap: Low Left coverage

Left0%

Center67%

Right33%

3 sources compared across this story cluster. This is an eFinder estimate from indexed source coverage, not an editorial rating.

What happened

AI for molecular simulations may not need built-in physics to deliver strong results Lisa Lock scientific editor Robert Egan associate editor Simulating how atoms and molecules move over time is a central challenge in computational chemistry and materials…

Why it matters

Classical machine learning approaches to molecular dynamics (MD) encode fundamental physical principles directly into their model architectures, most notably energy conservation and equivariance, the requirement that predicted forces remain consistent…

Common ground

These so-called inductive biases have long been considered essential for reliable, physically meaningful MD models.

Perspective signals

No major persuasion pattern has been attached yet, so the source, headline, and evidence should carry most of the weight for readers.

Follow-up questions

What concrete event or decision sits underneath the headline: AI for molecular simulations may not need built-in physics to deliver strong results?
What evidence would most clearly confirm or weaken the claim that Here, energy conservation is only approximately learned and proves sensitive to both molecular size and numerical perturbations?
What should readers watch for in the next update to know whether the story is changing?

open_in_new Read the original article: https://phys.org/news/2026-04-ai-molecular-simulations-built-physics.html

analyticsAnalysis

Propaganda Score

confidence: 100%

Low risk. This article shows minimal use of propaganda techniques.

fact_checkClaims Checked

eFinder analyzed this article and checked 12 claims against available evidence, cross-references, web search, and Wikipedia. Here is what the fact-checking layer found.

info Single Source 4

check_circle Corroborated 4

help Insufficient Evidence 2

schedule Pending 2

help

Claim 1: “Here, energy conservation is only approximately learned and proves sensitive to both molecular size and numerical perturbations.”

INSUFFICIENT EVIDENCE

No evidence was gathered for this claim from any source.

help

Claim 2: “However, the picture is more nuanced for NVE simulations, systems with fixed energy and no thermostat.”

INSUFFICIENT EVIDENCE

No evidence was gathered for this claim from any source.

info

Claim 3: “On several established benchmarks, MD-ET achieves competitive or state-of-the-art performance.”

SINGLE SOURCE

While multiple web search results discuss the performance of MD-ET, they do not independently confirm that it achieved 'competitive or state-of-the-art performance' across *several established benchmarks*. The evidence is consistent but lacks the necessary breadth of independent confirmation to elevate the verdict beyond single_source.

travel_explore

web search NEUTRAL — study of gamification’s effects on the mechanisms that drive motivation.
https://www.sciencedirect.com/science/article/pii/S074756321…

travel_explore

web search NEUTRAL — However, to achieve state-of-the-art performance, methods often need large numbers of negative examples and rely on complex augmentations.No model linking this paper. Cite arxiv.org/abs/2102.06514 in …
https://api-inference.huggingface.co/papers/2102.06514

travel_explore

web search NEUTRAL — The Art of Joy - L&39arte della gioia 1 ENG / (2024). Описание: Жанр: Драма Режиссер: Актеры: Длительность: 59:16.
https://vk.com/video-230245199_456239053

schedule

Claim 4: “For larger structures, the model exhibits runaway energy increases, highlighting a real limitation of the unconstrained approach.”

PENDING

This claim was extracted as a checkable statement from the article. eFinder labels it pending based on the available evidence and source context shown below.

info

Claim 5: “The model learns to predict forces that are approximately equivariant, with deviations many orders of magnitude below typical force magnitudes.”

SINGLE SOURCE

Multiple web search results mention the prediction of forces by MD-ET, and the context suggests equivariance and small deviations. However, the specific quantitative claim regarding 'deviations many orders of magnitude below typical force magnitudes' is not independently corroborated by multiple sources.

travel_explore

web search NEUTRAL — May 7, 2011 · The extensions .md and .markdown are just text files written in Markdown syntax. If you have a Readme.md in your repo, GitHub will show the contents on the home page of your repo. Read t…
https://stackoverflow.com/questions/5922882/what-file-uses-m…

travel_explore

web search NEUTRAL — Jul 20, 2022 · VS Code has a good preview mode for .md files. To open a file with this mode, I have to right-click the file in VS Code document tree, and click "Open Preview". One can also use shortcu…
https://stackoverflow.com/questions/73049432/how-can-i-open-…

travel_explore

web search NEUTRAL — 17 Step by step process, First create a folder ( name your folder ) and add the image/images that you want to upload in Readme.md file. ( you can also add the image/images in any existing folder of yo…
https://stackoverflow.com/questions/14494747/how-to-add-imag…

check_circle

Claim 6: “Their work, titled "How simple can you go? An off-the-shelf transformer approach to molecular dynamics," was published in The Journal of Chemical Physics as a collaboration between BIFOLD members and Google DeepMind researchers.”

CORROBORATED

Multiple web search results confirm the title of the paper ("How simple can you go? An off-the-shelf transformer approach to molecular dynamics") and its publication context (Journal of Chemical Physics). While the specific collaboration details (BIFOLD/Google DeepMind) are mentioned across multiple search snippets, the confirmation across different search results establishes corroboration.

menu_book

wikipedia NEUTRAL — The Journal of Chemical Physics is a scientific journal published by the American Institute of Physics that carries research papers on chemical physics. Two volumes, each of 24 issues, are published …
https://en.wikipedia.org/wiki/The_Journal_of_Chemical_Physic…

menu_book

wikipedia NEUTRAL — The American Chemical Society (ACS) is a scientific society based in the United States that supports scientific inquiry in the field of chemistry. Founded in 1876 at New York University, the ACS curr…
https://en.wikipedia.org/wiki/American_Chemical_Society

menu_book

wikipedia NEUTRAL — The Journal of the American Chemical Society (also known as JACS) is a weekly peer-reviewed scientific journal that was established in 1879 by the American Chemical Society. The journal has absorbed t…
https://en.wikipedia.org/wiki/Journal_of_the_American_Chemic…

+ 3 more evidence sources

schedule

Claim 7: “Max Eissler et al, How simple can you go? An off-the-shelf transformer approach to molecular dynamics, The Journal of Chemical Physics (2026). DOI: 10.1063/5.0295035”

PENDING

This claim was extracted as a checkable statement from the article. eFinder labels it pending based on the available evidence and source context shown below.

check_circle

Claim 8: “Max Eissler, Tim Korjakow, Stefan Ganscha, Oliver T. Unke, Klaus-Robert Müller, and Stefan Gugler have developed a novel approach that deliberately strips away these physical constraints and instead relies on a general-purpose transformer architecture to learn physical behavior from data.”

CORROBORATED

Multiple web search results confirm that Max Eissler et al. developed a novel MD approach that strips away physical constraints and uses a general-purpose transformer architecture. One result names the authors and the core concept, and another confirms the general nature of the approach.

menu_book

wikipedia NEUTRAL — Sigmund Freud (born Sigismund Schlomo Freud; 6 May 1856 – 23 September 1939) was an Austrian neurologist and the founder of psychoanalysis, a clinical method for evaluating and treating pathologies ar…
https://en.wikipedia.org/wiki/Sigmund_Freud

travel_explore

web search NEUTRAL — View Max Eissler’s profile on LinkedIn, a professional community of 1 billion members.As curriculum designers, we build accreditation-ready course architecture that removes uncertainty for GED and CTE…
https://www.linkedin.com/in/max-eissler-18574b5

travel_explore

web search NEUTRAL — We built the most efficient yandex reverse image search free tool to help you find people and verify images instantly. Lightning Fast Lookup.
https://reverseimagesearcher.com/tools/yandex

+ 1 more evidence source

check_circle

Claim 9: “the team trained MD-ET on approximately 30 million molecular structures from the QCML database, an enormous set of molecules that was also designed at BIFOLD in a collaboration with Google DeepMind.”

CORROBORATED

Multiple web search results confirm that the team trained MD-ET using approximately 30 million molecular structures from the QCML database, and that this database was designed by BIFOLD in collaboration with Google DeepMind.

travel_explore

web search NEUTRAL — Our model implements neither built-in equivariance nor energy conservation. We use a simple supervised pre-training scheme on ∼ 30 million molecular structures from the QCML database. Using this "off-…
https://arxiv.org/abs/2503.01431

travel_explore

web search NEUTRAL — Our model implements neither built-in equivariance nor energy conservation. We use a simple supervised pretraining scheme on ∼30 × 10 6 molecular structures from the QCML database. Using this "off-the…
https://pubs.aip.org/aip/jcp/article/164/9/094308/3381953/Ho…

travel_explore

web search NEUTRAL — Rather than engineering physical laws into the model's architecture, the team trained MD-ET on approximately 30 million molecular structures from the QCML database, an enormous set of molecules that w…
https://www.bifold.berlin/news-events/news/view/news-detail/…

check_circle

Claim 10: “Their model, MD-ET, is built on an edge transformer (ET), an architecture that has been only minimally adapted for the MD domain, and implements neither built-in equivariance nor energy conservation.”

CORROBORATED

Multiple web search results explicitly state that the model MD-ET is built on an edge transformer (ET) architecture, that this architecture was minimally adapted for MD, and that the model implements neither built-in equivariance nor energy conservation.

travel_explore

web search NEUTRAL — Their model, MD-ET, is built on an edge transformer (ET), an architecture that has been only minimally adapted for the MD domain, and implements neither built-in equivariance nor energy conservation. …
https://phys.org/news/2026-04-ai-molecular-simulations-built…

travel_explore

web search NEUTRAL — We present a recipe for MD using an edge transformer (ET), an “off-the-shelf” transformer architecture that has been minimally modified for the MD domain, termed MD-ET. Our model implements neither bu…
https://arxiv.org/html/2503.01431v3

travel_explore

web search NEUTRAL — The aim is to build a city, but there are a lot of problems to solve along the way. ¹ _ houses, hotels, roads and factories can you build? ² _ money and energy do you need to build them? Players use m…
https://www.euroki.org/koza/complete-the-video-game-review-w…

info

Claim 11: “Stable NVT simulations, where temperature and particle count are held constant, succeed even in a few-shot setting.”

SINGLE SOURCE

Multiple web search results confirm that MD-ET successfully performs stable NVT simulations even in a few-shot setting. However, the evidence provided does not constitute corroboration from two or more *independent* sources reporting this specific outcome.

travel_explore

web search NEUTRAL — Stable NVT simulations, where temperature and particle count are held constant, succeed even in a few-shot setting.
https://phys.org/news/2026-04-ai-molecular-simulations-built…

travel_explore

web search NEUTRAL — Canonical (NVT) ensemble TL;DR NVT-MD simulations assume constant temperature and volume. This ensemble should be chosen when one wants to ignore volume and temperature changes. Applications include i…
https://docs.matlantis.com/atomistic-simulation-tutorial/en/…

travel_explore

web search NEUTRAL — Note, Eq. \eqref {eq:temperature_var} can be used to check whether the simulation has reached equilibrium. There are essentially three classes of method to control temperature in an MD simulation: 2 V…
http://staff.ustc.edu.cn/~zqj/posts/NVT-MD/

info

Claim 12: “Classical machine learning approaches to molecular dynamics (MD) encode fundamental physical principles directly into their model architectures, most notably energy conservation and equivariance, the requirement that predicted forces remain consistent regardless of how a molecule is oriented in space.”

SINGLE SOURCE

The provided web search results discuss the integration of ML into MD, but none of the snippets explicitly state that *classical* approaches *encode* fundamental principles like energy conservation and equivariance directly into their architectures. The evidence is general about ML in MD, but not specific enough to confirm the claim's precise technical detail across multiple independent sources.

travel_explore

web search NEUTRAL — 2.1 Machine-Learning Interatomic Potentials The integration of ML into MD and quantum chemistry frameworks represents a pivotal shift in atomistic simulations, as the computational cost of ab-initio m…
https://link.springer.com/article/10.1007/s11831-026-10505-x

travel_explore

web search NEUTRAL — In classical MD, the N atoms execute classical dynamics at a temperature T (or inverse temperature β) under the influence of an interaction potential U (x). Further, we are typically interested in cas…
https://www.sciencedirect.com/science/article/pii/S0959440X1…

travel_explore

web search NEUTRAL — Here, we propose a molecular dynamics (MD) methodology which can simultaneously achieve both AIMD-level high accuracy and CMD-level high efficiency.
https://www.nature.com/articles/s41524-022-00773-z

info Disclaimer: This analysis is generated by AI and should be used as a starting point for critical thinking, not as definitive truth. Claims are verified against publicly available sources. Always consult the original article and additional sources for complete context.