Neural Petri Flows: Hard-Wiring Physical Conservation Laws into Chemical Reaction AI
結合 Petri 網與深度學習的神經 Petri 流（NPF）：將物理守恆定律寫入化學反應預測
Traditional AI models struggle to guarantee that predicted chemical reactions adhere to physical laws. Researchers introduced Neural Petri Flows (NPF), mapping chemistry to Petri nets where places represent bonds/valence and tokens represent bond orders. By hard-wiring physical conservation and non-negativity into parameter-free layers, NPF only needs to learn reaction rate laws. Without training, NPF outperforms RXNMapper in atom mapping. When trained on USPTO-480K, it achieves highly accurate forward prediction while guaranteeing that every predicted output is a physically valid molecule without post-filtering.
Key points
Hard-Wired Physics Laws
NPF hard-wires valence budgets and bond conservation into parameter-free layers, guaranteeing physically valid outputs regardless of model weights.
Zero-Shot Atom Mapping
Without any training, NPF achieves 88.8% on the Golden set and 88.7% on EnzymeMap, outperforming established baselines like RXNMapper.
Exceptional Data Efficiency
When trained on just a 1% subset of the USPTO-480K dataset, NPF still correctly predicts 67.4% of the chemical reaction products.
Guaranteed Molecule Validity
Using electrons as tokens, NPF achieves 90.5% accuracy on FlowER elementary steps, with every top-1 prediction guaranteed to be a valid molecule without any filtering.
How it works
Why it matters
Traditional chemical AI models often hallucinate physically impossible molecules, requiring heavy filtering. NPF proves the power of hard-wiring exact physical inductive biases. By letting the neural network only learn the "rate laws" (reaction propensities) while locking down the conservation laws, it reduces data hunger dramatically (working on just 1% data) and guarantees structural validity. This is a game-changer for high-precision chemical synthesis planning and drug discovery.
Who it affects
- AI Developer
- AI Researcher
- Enterprise Leader
How to use it
- 1Atom Mapping: Tracking atom transitions in complex and enzymatic chemical reactions
- 2Forward Reaction Prediction: Predicting the final product molecules from given reactants
- 3Reaction Classification: Automatically classifying reaction types (e.g., EC numbers in ECREACT)
Limitations & caveats
- Relies heavily on the exact mapping between Petri nets and chemical structures; systems with complex transition-metal coordination or non-classical bonding remain challenging to model.
Related
How Conformal Prediction Sets Quantify Information Gain: An Information-Theoretic Foundation
符合性預測集合如何量化資訊增益：資訊理論的新視角
This study establishes an information-theoretic foundation for using Conformal Prediction set sizes as uncertainty metrics, linking them to Shannon mutual information.
4D-HOF: Feed-Forward 4D Hand-Object Interaction Reconstruction via Flow Matching
4D-HOF：利用流匹配技術實現前饋式 4D 手部與物體互動重建
4D-HOF is a feed-forward framework that uses conditional flow matching to refine coarse initial hand-object states into physically and geometrically consistent 4D reconstructions.
IdeaAnchor: Teaching LLMs to Turn Literature into Research Ideas
IdeaAnchor：教導大語言模型將學術文獻轉化為研究點子
Researchers developed IdeaAnchor, a paradigm that trains LLMs to generate high-quality research ideas by leveraging structured specifications mined from published papers.