How the RFT Engine Computes Molecular Physics

A first-principles engine that predicts intermolecular interaction energies directly from geometry, with zero fitted parameters.

What it computes

The engine predicts intermolecular interaction energies, that is, how strongly two molecules attract or repel one another, directly from atomic geometry. This is the same quantity that rigorous quantum chemistry computes with methods such as CCSD(T) and SAPT, but it is obtained without an iterative wavefunction calculation.

The physical picture

The total interaction is assembled from the well established physical components that govern how molecules interact:

• Dispersion, the van der Waals attraction between electron clouds.
• Exchange and Pauli repulsion, the cost of overlapping closed electron shells.
• Electrostatics, the interaction between the molecules' charge distributions.
• Induction, the polarization of each molecule in the field of the other.

These are the same components that a rigorous quantum-chemistry energy decomposition identifies. The engine evaluates each one explicitly.

What is different, and why it generalizes

Conventional tools take one of two routes. They either solve the electronic Schrodinger equation iteratively, which is accurate but slow, or they train a neural network on millions of reference calculations, which is fast but bounded by the training data. The RFT engine does neither.

Each component is a closed-form expression whose constants are fixed by the geometry of the underlying framework, not fitted to any dataset. Because there are no adjustable parameters, there is nothing to overfit. The same constants apply to every molecule, every element, and every geometry.

Why it appears to work

When the constants come from the structure of the framework rather than from data, the model is not interpolating between examples. It is computing physics. That is why its accuracy on molecules it has never been shown matches its accuracy on any other, a property usually described as in-sample accuracy equalling out-of-sample accuracy. It is also why the engine reaches the accuracy tier of methods trained on millions of calculations while using none.

How it is validated

The engine is checked against gold-standard quantum chemistry, including symmetry-adapted perturbation theory (SAPT2+), near full-configuration-interaction references, and CCSD(T) at the complete-basis-set limit, on standard benchmarks such as S66, A24, and PLA15. It reproduces the reference energies with zero fitted parameters. On the PLA15 protein-ligand set it matches or beats trained machine-learning potentials that each required millions of quantum-chemistry calculations. The full results are given in Validation Against Gold-Standard Quantum Chemistry.

Why it is fast

Each component is a direct geometric evaluation rather than an iterative solve or a large neural-network inference, so a full calculation runs in milliseconds. This is fast enough to run inside a web browser, at millions of systems per second on a consumer GPU.

What remains proprietary

The specific mathematical framework that fixes the constants in these expressions is not disclosed. This note describes what the engine computes and why the approach generalizes, not how the framework is constructed.