QuantaMind: Simulating Full Chemical Reactions on a Single GPU, from Structure Prediction to Watching Reactions Unfold

Forum topic · 小凯 · 2026-09-18

Summary

QuantaMind, a reactive machine learning force field developed by Shanghai-based Molecule Heart (Fenzixin), published in Science Advances, claims to tackle the long-standing 'impossible triangle' of computational chemistry: accuracy approaching first-principles/DFT-level electronic structure, speed sufficient for long trajectories, and scale reaching ~100,000 atoms. The platform reports inference speeds of 4.2×10⁻⁶ s/atom/step on a single NVIDIA A100 80GB GPU (later improved to 2.1×10⁻⁶ in extended tests), reducing the per-step cost of a 100,000-atom reactive simulation to about 0.25 seconds, and was validated on a full 17,792-atom protein system. Unlike conventional non-reactive machine learning force fields with fixed molecular topology, QuantaMind can track proton transfer, bond breaking and formation, and reaction intermediates—from small molecules to enzymes. The team also demonstrated femtosecond-resolution animations of catalytic reactions, letting researchers visually inspect reaction mechanisms rather than checking coordinate data frame by frame. The article notes open questions: no independent replication yet, no published error distributions versus DFT, uncertainty about whether 100-nanosecond timescales and 100,000-atom scales hold simultaneously, and a remaining gap of several orders of magnitude to real cellular environments.

QuantaMind: One GPU, One Atomic Movie — From Predicting Structures to Seeing Reactions

A molecular structure diagram looks like a still photo. Real molecules are not still. In solution they constantly collide, rotate, break and reform chemical bonds, and shuttle protons back and forth between atoms.

A single bond-breaking event happens on the femtosecond scale—a quadrillionth of a second—while a complete enzyme-catalyzed reaction can last milliseconds. To see every atomic step in between, a simulation must span roughly twelve orders of magnitude in time. Two more hurdles remain: accuracy close to first-principles calculations, and scale large enough to hold a biological system. QuantaMind, from Shanghai-based Molecule Heart (分子之心), recently published results in *Science Advances* attempting to push all three at once.

The trade-off that couldn't be won

Computational chemistry's difficulty is a three-way trade-off:

Accuracy: quantum-mechanical descriptions, since bond breaking and formation are fundamentally electron rearrangements that classical spring-like force fields cannot capture.
Speed: long enough trajectories to see a reaction's full path.
Scale: tens of thousands to hundreds of thousands of atoms to contain a piece of cell membrane or a complete catalytic active site.

For decades, reactive molecular dynamics was confined to tiny systems and very short time windows. The scenarios that actually need it—how a drug molecule chemically transforms in an enzyme's active site, or how a proton hops along a hydrogen-bonded chain—were out of reach.

The three key numbers

Inference speed: 4.2×10⁻⁶ s/atom/step on a single NVIDIA A100 80GB GPU, claimed to match state-of-the-art machine learning force fields; extended tests pushed this to 2.1×10⁻⁶ s/atom/step.
Derived figure: for a ~100,000-atom reactive system, a single simulation step takes ~0.25 seconds on one card.
Validation: efficiency demonstrated on a full 17,792-atom protein system—smaller by an order of magnitude than 100,000 atoms, but a real biomolecule rather than a fragment.

One caveat: the number of timesteps for a 100,000-atom run depends on the timestep, which was not officially stated. If a conventional 1 fs timestep is assumed, 100 ns would require 10⁸ steps—over half a year on a single card. This is inference, not a refutation; it simply means the wall-clock time behind the "hundred-nanosecond" claim cannot be derived from public information.

| Metric | Paper | Extended tests | |---|---|---| | Inference speed | 4.2×10⁻⁶ s/atom/step | 2.1×10⁻⁶ s/atom/step | | Validation system | 17,792-atom full protein | ~100,000-atom reactive system | | Time scale | Tens of ns | Hundreds of ns | | Per-step cost | Not disclosed | ~0.25 s (100k atoms) | | Hardware | Single NVIDIA A100 80GB | Single A100 80GB |

What "reactive" costs

Recent machine learning force fields can compute energies and run large systems quickly, but share one weakness: they are non-reactive. Atomic connectivity is fixed in the topology; when bonds must break or form, the model has no description for it.

QuantaMind is positioned as a reactive ML force field. It can track proton transfer, chemical bond formation and breaking, and reaction intermediates, covering systems from small molecules and drug candidates up to proteins and enzymes.

This distinction matters concretely in drug development. Why two visually similar molecules show different binding and reaction behavior, or how a single proton transfer changes the outcome, is a black box in non-reactive models. Mechanism tracking means researchers can see the "why" computationally, not just a "can/cannot" score.

Rendering trajectories as watchable films

The real usability test is visualization. In one demo, a catalytic reaction—hydrogen molecules adsorbing onto a catalyst surface, splitting into atoms, and desorbing as water—was rendered as an animation with femtosecond resolution.

For computational chemists, the value is in verification cost: reaction paths that once required frame-by-frame inspection of coordinates and energies can now be watched directly, making anomalous frames easier to spot. A side benefit: experimentalists who don't run simulations can understand the trajectories too—a real communication saving in interdisciplinary teams.

The competition

The technical route—training ML models on large-scale quantum chemistry data, then replacing expensive first-principles calculations with inference—includes Meta's UMA, Microsoft's MatterSim, and Google DeepMind's GEMS. The differentiator lies in reactive coverage and system types.

Molecule Heart founder Xu Jinbo developed the first effective AI protein-structure-prediction algorithm in 2016, later credited with inspiring the first generation of AlphaFold. Founded in 2022, the company completed a Series A round totaling over US$100 million in June 2026 and opened its bio-economy operating system, MoleculeOS, in July. QuantaMind completes the matrix: MoleculeOS previously covered generation–design–prediction; QuantaMind adds mechanistic understanding.

Four things not yet verified

1. Independent replication — none has appeared yet. 2. "Near-DFT accuracy" — ML force fields usually perform well within the training distribution; the out-of-distribution error growth rate, critical for real R&D use, has no published distribution numbers. 3. Whether hundred-nanosecond and 100,000-atom scales hold simultaneously. 4. Real cellular environments — a cell contains billions of atoms; 100,000 atoms covers roughly a membrane patch or one active site. Several orders of magnitude remain, with nonlinearly rising difficulty in training-data coverage at each step.

What happens when the barrier drops

The end scenario is concrete: an ordinary lab that once queued for supercomputer time and needed a computational-chemistry postdoc could produce usable 100,000-atom results on a workstation with a single A100.

What changes is not one team's ceiling, but *who is qualified to ask the question*. Questions once reserved for compute-rich institutions become accessible to small and mid-size teams.

And when simulation moves from "predicting outcomes" to "reconstructing processes," will researchers spend more time designing experiments—or verifying the mechanisms AI hands them?

References

1. Molecule Heart QuantaMind research in *Science Advances*, 2026-09 — Science and Technology Daily: https://www.stdaily.com/web/gdxw/2026-09/17/content_583085.html 2. Science and Technology Daily, 2026-09-16: https://www.stdaily.com/web/gdxw/2026-09/16/content_582193.html 3. QbitAI coverage, 2026-09-15, via BAAI Link: https://link.baai.ac.cn/@liangzw/117280329474008819 4. Sohu Tech, 2026-09: https://www.sohu.com/a/1076656175_121421892 5. 6th "Haiju Yingcai" Global Innovation & Entrepreneurship Competition AI4S track, 2026-09-17: https://news.qq.com/rain/a/20260917A05Q9N00