Neural Potentials and Score-Based TI

• Neural Potentials and Score-Based TI are advanced methods that combine neural network-parameterized Hamiltonians with score-based diffusion models for high-dimensional free-energy estimation.
• They use a continuous alchemical path and time-dependent potentials to efficiently sample intermediate ensembles, eliminating the need for multiple MCMC or MD simulations.
• The approach demonstrates scalability and accuracy through rigorous free-energy calculations validated against Monte Carlo benchmarks in complex systems.

Neural Potentials and Score-Based Thermodynamic Integration (TI) refer to the integration of neural network–parameterized Hamiltonians and energy-based denoising diffusion models for efficient, high-dimensional free-energy estimation. These approaches enable the estimation of free-energy differences by learning a time-dependent potential along a continuous “alchemical” path, sampled via score-based diffusion, thereby sidestepping the traditional need for multiple Markov Chain Monte Carlo (MCMC) or molecular dynamics (MD) simulations at a sequence of intermediate states. Neural TI provides rigorous, scalable, and accurate free-energy calculations for complex systems, particularly in statistical physics and molecular modeling (Máté et al., 2024).

1. Thermodynamic Integration Fundamentals

Thermodynamic Integration (TI) estimates the free-energy difference $\Delta F$ between two systems by integrating over a parametric pathway in Hamiltonian space. Given microscopic coordinates $x$ and a one-parameter family of Hamiltonians $H(x,t)$ with $t \in [0,1]$ interpolating between the reference ($H_0$) at $t=0$ and the target ($H_1$) at $t=1$,

$Z(t) = \int \mathrm{d}x\, \exp[-\beta H(x,t)]$

$p_t(x) = \frac{1}{Z(t)} \exp[-\beta H(x,t)]$

The canonical free-energy change is

$x$0

where $x$1 denotes the Boltzmann average with respect to $x$2, and $x$3 is the inverse temperature. This classical approach requires sampling from each $x$4, a computationally demanding requirement for high-dimensional or complex systems (Máté et al., 2024).

2. Neural Network Parameterization of Time-Dependent Potentials

A principal innovation is to replace hand-crafted $x$5 with a trainable neural Hamiltonian $x$6, with $x$7 as the neural potential:

$x$8

where $x$9 is the analytic kinetic term and $H(x,t)$0 is a neural network constrained such that $H(x,t)$1 and $H(x,t)$2. In practice, a soft-core Lennard-Jones (LJ) form is used to regularize singularities,

$H(x,t)$3

The neural potential $H(x,t)$4 is implemented as an equivariant graph network, receiving the full configuration $H(x,t)$5 and continuous $H(x,t)$6, with time injected via learned MLP embeddings. These architecture choices ensure $H(x,t)$7-equivariance and appropriate boundary behavior at $H(x,t)$8, thus enabling a smooth, data-driven interpolation between ensembles (Máté et al., 2024).

3. Score-Based Diffusion Model for Intermediate Sampling

Sampling from each $H(x,t)$9 is performed using a continuous-time, score-based diffusion model. The core is the score function

$t \in [0,1]$0

which serves both as the gradient of the learned energy model and as an approximate score for the evolving density $t \in [0,1]$1. The forward process follows an Itô SDE that gradually adds noise; the reverse-time SDE, utilizing $t \in [0,1]$2, transports samples from a tractable reference (e.g., ideal gas) to any intermediate or final $t \in [0,1]$3.

Intermediate ensemble sampling proceeds by integrating the reverse SDE from $t \in [0,1]$4 to any $t \in [0,1]$5, or via a probability flow ODE. This enables efficient, direct sampling at arbitrary $t \in [0,1]$6, thereby eliminating the need to run separate simulations at multiple coupling strengths—a key limitation in standard TI methods (Máté et al., 2024).

4. Training via Score Matching

The denoising-score matching objective trains the network to learn $t \in [0,1]$7 by predicting the noise added in the forward diffusion process:

$t \in [0,1]$8

$t \in [0,1]$9

$H_0$0

Here, $H_0$1 triples are sampled via reference data, a uniformly random $H_0$2, and i.i.d. Gaussian noise. Training proceeds by minimizing $H_0$3 through standard backpropagation and stochastic gradient descent, updating the neural potential's parameters until the score estimates drive accurate diffusion-based sampling (Máté et al., 2024).

5. Free Energy Estimation from a Single Network

After training, free-energy differences are estimated along the learned neural path:

$H_0$4

In practice, time points $H_0$5 are selected, with $H_0$6 samples $H_0$7 generated via score-based dynamics. The estimator is

$H_0$8

where $H_0$9 are quadrature weights. Since the $t=0$0 ideal-gas partition function $t=0$1 is analytically known,

$t=0$2

Ensembles at all $t=0$3 are available from a single trained model, supporting evaluation of both canonical and grand-canonical partition functions, e.g., for direct calculation of excess chemical potentials $t=0$4 from $t=0$5 (Máté et al., 2024).

6. Empirical Validation and Results

Neural TI has been validated on 3D Lennard-Jones fluids within periodic boxes over a range of densities ($t=0$6, $t=0$7 to $t=0$8). Key findings include:

• Radial distribution functions $t=0$9 sampled from the trained score-based model match those from Monte Carlo references across the gas–liquid transition.
• Grand-canonical particle-number distributions $H_1$0 and excess chemical potentials $H_1$1 inferred via the TI approach closely track values from grand-canonical Monte Carlo.
• Canonical free-energy differences up to $H_1$2 (corresponding to coupling of up to $H_1$3 degrees of freedom) are accurately estimated from a single neural network–based sampling process.
• The framework demonstrates strong scaling to high-dimensional systems, with a single diffusion model covering all coupling strengths (Máté et al., 2024).

7. Strengths, Limitations, and Prospects

Strengths of Neural TI include its elimination of multiple intermediate $H_1$4-windows, data-driven adaptation to the optimal alchemical path, and tractability for hundreds of degrees of freedom at once. Limitations remain in model capacity, as the expressivity of $H_1$5 and the quality of SDE/ODE integration affect accuracy, especially in rough or rare-event–dominated landscapes. Computational cost is front-loaded in network training and requires careful architectural design (e.g., equivariance, explicit time dependence).

Future extensions include application to multi-component liquids, biomolecular transformations, or ab-initio potentials via integration with neural force fields (SchNet, MACE, E(3)-GNN). Further algorithmic improvements may involve adaptive quadrature, variance reduction by control variates, and constraints for other statistical ensembles (e.g., NPT) (Máté et al., 2024).

Neural TI synthesizes advances in energy-based modeling, diffusion generative dynamics, and statistical mechanics, offering a unified framework for large-scale, rigorous free-energy estimation via neural potentials and score-based sampling.