Force Matching for Coarse-Grained Force Fields

• Force Matching (FM) is a projection-based approach that constructs coarse-grained force fields by minimizing the mean-squared error between reference and predicted forces.
• It employs conditional expectations and connects with thermodynamic integration, relative entropy minimization, and advanced mean force matching techniques for machine-learned potentials.
• Practical applications involve mapping atomistic trajectories to coarse-grained variables using least-squares optimization to recover equilibrium properties in lower-dimensional representations.

Force Matching (FM) is a statistical, projection-based approach for constructing coarse-grained (CG) force fields by minimizing the mean-squared deviation between reference atomic or fine-grained forces and forces predicted by a candidate CG model. Initially formulated for molecular coarse-graining, FM has evolved into a general framework grounded in conditional expectations, Euler–Lagrange equations, and information metrics. The method provides a principled route to derive CG potentials that recover equilibrium properties of high-dimensional atomistic systems in lower-dimensional representations. FM is closely connected with thermodynamic integration, relative entropy minimization, and, in its recent variants, mean force matching (MFM) for machine-learned potentials.

1. Mathematical and Probabilistic Framework

The modern formulation of FM is as an $L^2$ projection in the probability space defined by the atomistic system's Boltzmann distribution $\mu(dx) = Z^{-1} e^{-\beta U(x)} dx$ with atomistic coordinates $x \in \mathbb{R}^{3N}$ and energy $U(x)$ (Kalligiannaki et al., 2015). Given a CG map $\xi:\mathbb{R}^{3N} \rightarrow \mathbb{R}^m$, FM seeks a CG force field $G(z)$ that best approximates a reference (often atomistic) local force $h(x)$ in mean-square:

$$\min_{G \in L^2(\mu;\xi)} \int \| h(x) - G(\xi(x)) \|^2 \, \mu(dx).$$

This problem admits a unique solution:

$G^*(z) = E_\mu[h \mid \xi = z],$

where $E_\mu[\cdot \mid \xi = z]$ is the conditional expectation, i.e., the orthogonal projection of $\mu(dx) = Z^{-1} e^{-\beta U(x)} dx$0 onto the subspace of observables measurable with respect to $\mu(dx) = Z^{-1} e^{-\beta U(x)} dx$1.

This perspective recasts FM as a projection theorem in Hilbert spaces, underpinning its statistical optimality and generalizing to both linear and nonlinear CG maps.

2. Connection with Thermodynamic Integration and Mean-Force Estimation

FM can be linked directly to thermodynamic integration (TI), providing a formula for the local mean force associated with the potential of mean force (PMF) $\mu(dx) = Z^{-1} e^{-\beta U(x)} dx$2:

$\mu(dx) = Z^{-1} e^{-\beta U(x)} dx$3

The TI formalism yields an explicit estimator for the PMF gradient at $\mu(dx) = Z^{-1} e^{-\beta U(x)} dx$4 as a conditional expectation:

$\mu(dx) = Z^{-1} e^{-\beta U(x)} dx$5

with

$\mu(dx) = Z^{-1} e^{-\beta U(x)} dx$6

where $\mu(dx) = Z^{-1} e^{-\beta U(x)} dx$7, $\mu(dx) = Z^{-1} e^{-\beta U(x)} dx$8 is a weight matrix, and $\mu(dx) = Z^{-1} e^{-\beta U(x)} dx$9 with $x \in \mathbb{R}^{3N}$0 (Kalligiannaki et al., 2015). For practical CG mappings, this reduces to well-defined, computationally tractable mean-force estimators.

In the context of machine-learned CG potentials, mean force matching (MFM) leverages constrained molecular dynamics to provide direct low-variance estimates of $x \in \mathbb{R}^{3N}$1 at discrete $x \in \mathbb{R}^{3N}$2, defining an objective

$x \in \mathbb{R}^{3N}$3

where $x \in \mathbb{R}^{3N}$4 is the mean force at $x \in \mathbb{R}^{3N}$5, $x \in \mathbb{R}^{3N}$6 is the model force, and $x \in \mathbb{R}^{3N}$7 indexes model parameters (Park et al., 16 Feb 2026).

3. Implementation and Algorithmic Procedures

In practice, FM proceeds by generating atomistic reference trajectories, mapping atomic forces to pre-defined CG degrees of freedom, and solving a linear least-squares problem to fit force-field parameters (Alvares et al., 2023, Park et al., 16 Feb 2026). For tabulated spline-based CG potentials $x \in \mathbb{R}^{3N}$8, the objective for each configuration $x \in \mathbb{R}^{3N}$9 and bead $U(x)$0 is

$U(x)$1

where $U(x)$2 is the sum of atomistic forces for bead $U(x)$3, mapped for each configuration (Alvares et al., 2023). Software implementations such as VOTCA-cg automate this process for various molecular systems.

For MFM, the required mean forces $U(x)$4 are estimated by averaging instantaneous projected forces over constrained MD trajectories at fixed $U(x)$5. This label denoising enables the scaling of highly expressive ML architectures (e.g., MACE, eSEN), significantly reduces data requirements, and improves statistical convergence (Park et al., 16 Feb 2026).

4. Theoretical Properties, Relative Entropy, and Equivalences

FM is equivalent, up to constants and higher-order terms, to enforcing orthogonality between the candidate CG force and intrinsic force fluctuations eliminated by coarse-graining. More formally, minimizing the FM loss is asymptotically equivalent to minimizing the $U(x)$6-distance (squared gradient norm) between the CG potential and the true PMF, while relative entropy minimization aligns their $U(x)$7-distance (potential values). For a model potential $U(x)$8, the relative entropy cost is

$U(x)$9

while FM yields

$\xi:\mathbb{R}^{3N} \rightarrow \mathbb{R}^m$0

with the best approximations differing by a constant (Kalligiannaki et al., 2015). This suggests that FM and relative entropy minimization are effectively equivalent in fitting the PMF up to constant shift, but FM focuses on gradients and forces.

5. Performance, Benchmarks, and Empirical Observations

FM-derived CG force fields have been benchmarked in physical and biochemical contexts. In ZIF-8, FM-based potentials reproduce atomistic structures, lattice constants (within $\xi:\mathbb{R}^{3N} \rightarrow \mathbb{R}^m$1 Å), and certain elastic constants, outperforming generic CG models like MARTINI in structure preservation and transfer properties (Alvares et al., 2023). However, mechanical property predictions remain sensitive to post-hoc corrections, such as volume-dependent potential terms, and spurious negative elastic constants may arise if not properly constrained.

In biomolecular applications, MFM has demonstrated marked improvements in thermodynamic consistency, requiring $\xi:\mathbb{R}^{3N} \rightarrow \mathbb{R}^m$2 fewer training samples and $\xi:\mathbb{R}^{3N} \rightarrow \mathbb{R}^m$3 less simulation time than FM to achieve equivalent accuracy on held-out proteins. This data-efficiency is attributed to removal of irreducible noise in the FM labels. MFM enables training of state-of-the-art neural potentials (e.g., MACE $\xi:\mathbb{R}^{3N} \rightarrow \mathbb{R}^m$4 parameters; eSEN $\xi:\mathbb{R}^{3N} \rightarrow \mathbb{R}^m$5 parameters) (Park et al., 16 Feb 2026). The table below summarizes select benchmarking results for CG protein force fields:

Model	Test MSE FM	Test MSE MFM	Training Points Ratio
SchNet	40.1	32.7	375× fewer for MFM
MACE	34.6	26.4	375× fewer for MFM
eSEN	24.4	19.2	375× fewer for MFM

Additional studies report FM's ability to capture phase transitions such as the swing effect in porous solids, outperforming alternative approaches in reproducing subtle collective behaviors upon guest loading (Alvares et al., 2023).

6. Practical Considerations and Limitations

Key practical elements for successful FM implementation include:

• Mapping and Basis Functions: Appropriate CG mapping strategies, centering of beads, and use of spline-tabled potentials are crucial for accuracy and stability (Alvares et al., 2023).
• Sampling and Regularization: High-quality atomistic reference runs, often at elevated temperatures for crystalline solids to ensure phase-space coverage, are necessary. Regularization, pressure- or virial-matching corrections may be required for stability and transferability but can introduce sensitivity in thermomechanical predictions.
• Software and Automation: Packages such as VOTCA-cg, OpenMM for MD, and PyTorch or JAX for MLIP development are frequently used (Alvares et al., 2023, Park et al., 16 Feb 2026).
• Data-Efficiency: FM suffers from high label noise when degrees of freedom are heavily reduced, requiring large numbers of samples or blocks; MFM mitigates this via averaged mean-force labels (Park et al., 16 Feb 2026).
• Model Transfer and Generalization: MFM-trained models generalize better to unseen or out-of-distribution domains and support stable, large-scale model optimization (Park et al., 16 Feb 2026).

Limitations remain in the extrapolation of FM potentials beyond sampled regions, handling of non-analytic tabulated potentials, and in the design of pressure/volume corrections. Poorly regularized FM models can yield unphysical elastic constants or structural artifacts, especially in crystalline solids (Alvares et al., 2023).

7. Extensions, Variants, and Related Methodologies

Extensions of FM include:

• Generalized FM: FM is applicable to both linear and nonlinear CG mappings, as rigorously shown in the projection formalism, providing estimators for arbitrary observables such as end-to-end distances or bending angles (Kalligiannaki et al., 2015).
• Mean Force Matching (MFM): A variant of FM leveraging mean-force estimators via constrained dynamics, enabling low-variance training for large ML potentials and scaling to high accuracy and transferability (Park et al., 16 Feb 2026).
• Connection to Score Matching and Relative Entropy: Score matching targets the log-density gradient (the score) via noisy Laplacians and is generally less scalable and more restrictive in required data distributions than FM/MFM (Park et al., 16 Feb 2026).
• Applications in Quantum Chromodynamics (QCD): FM/matching concepts are also foundational in constructing effective Lorentz-force operators via one-loop matching in thermal QCD, with explicit renormalization and physical implications for heavy-quark diffusion (Laine, 2021).

The FM framework thus unifies variational, information-theoretic, and projection-based criteria for bottom-up coarse-graining, providing mathematical consistency, physical transparency, and extensibility to machine-learned and domain-adapted models.