Machine-Learned Many-Body Potentials for Charged Colloids reveal Gas-Liquid Spinodal Instabilities only in the strong-coupling regime of Primitive Models

Past experimental observations of gas-liquid and gas-crystal coexistence in low-salinity suspensions of highly charged colloids have suggested the existence of like charge attraction. Evidence for this phenomenon was also observed in primitive-model simulations of (asymmetric) electrolytes and of low-charge nanoparticle dispersions. These results from low-valency simulations have often been extrapolated to experimental parameter regimes of high colloid valency where like-charge attraction between colloids has been reported. However, direct simulations of highly charged colloids remain computationally demanding. To circumvent slow equilibration, we employ a machine-learning (ML) framework to construct ML potentials that accurately describe the effective colloid interactions. Our ML potentials enable fast simulations of dispersions and successfully reproduce the gas-liquid and gas-solid phase separation observed in primitive-model simulations at low charge numbers. Extending the ML-based simulations to higher valencies, where primitive-model simulations become prohibitively slow, also reveals like-charge attractions and gas-liquid spinodal instabilities, however only in the regime of strongly coupled electrostatic interactions and not in the weakly coupled Poisson-Boltzmann regime of the experimental observations of colloidal like-charge attractions.

💡 Research Summary

In this work the authors address the long‑standing controversy over the apparent attraction between like‑charged colloids observed in low‑salt, highly charged suspensions. Classical DLVO theory, based on a linearized Poisson‑Boltzmann (PB) description, predicts only screened‑Coulomb repulsion, yet experiments have reported gas‑liquid coexistence, clustering, void formation and unusually stable crystals that imply an effective attraction. Previous primitive‑model (PM) simulations, which treat ions explicitly while modelling the solvent as a dielectric continuum, have reproduced some of these anomalies but were limited to relatively low colloidal valences (Z≈10–20) because the number of counter‑ions required for charge neutrality grows dramatically with Z, making the simulations prohibitively slow.

To overcome this bottleneck the authors develop a machine‑learning (ML) framework that learns an effective many‑body colloid‑only potential from forces obtained in short PM molecular‑dynamics runs. They generate a training set consisting of 240 configurations spanning packing fractions η=0.001–0.45 for a fixed system of N=32 colloids (plus the appropriate number of monovalent counter‑ and co‑ions). For each configuration the ion‑averaged force on every colloid, F_PM, is recorded. A linear‑regression force‑matching scheme using Behler‑Parrinello‑type symmetry functions is then employed; the weights are optimized to minimize the root‑mean‑square error (RMSE≈0.02 k_BT/σ) and achieve a coefficient of determination R²≈0.98, indicating that the ML potential reproduces the PM forces with near‑perfect fidelity.

First, the authors validate the ML potential in the low‑valence regime (Z=10–20). Large‑scale simulations using the learned potential reproduce the gas‑liquid and gas‑solid phase separations previously reported from direct PM simulations, confirming that the ML model captures the essential many‑body ion‑mediated correlations.

Next, they apply the same potential to higher valences (Z=40, 60, 80) and explore a range of reduced temperatures σ/λ_B (where λ_B is the Bjerrum length). Two distinct regimes emerge:

1. Strong‑coupling regime (σ/λ_B ≲ 1) – Here the Bjerrum length is comparable to or larger than the colloid diameter, corresponding to low temperature, low dielectric constant, or high ion valence. Counter‑ions condense into a quasi‑two‑dimensional correlated layer on the colloid surface. The ML potential exhibits a short‑range attractive well that becomes deeper with increasing Z, leading to a clear gas‑liquid spinodal. The critical temperature rises with Z, and phase separation is observed at moderate packing fractions (η≈0.1–0.2).

2. Weak‑coupling (PB) regime (σ/λ_B ≫ 1) – Even for large Z, if the temperature is high enough (or the solvent dielectric constant large) the Bjerrum length is small, and the system behaves according to linear PB theory. In this regime the effective potential is purely repulsive; no spinodal or gas‑liquid instability is found. Thus the “volume‑term” attraction predicted by some PB‑based theories does not manifest in the simulations.

These findings lead to two major conclusions. First, the attractive interactions seen in low‑valence PM studies are not universally transferable to the high‑valence experimental systems; they only survive when the electrostatic coupling is strong enough to generate ion‑ion correlations beyond mean‑field. Second, the experimentally reported like‑charge attractions in low‑salt, highly charged colloids can only be explained by strong‑coupling physics, not by the weak‑coupling PB volume‑term mechanism.

The paper also discusses limitations. The training data are limited to N=32 colloids and a specific size ratio (σ_i = σ/20), so finite‑size and long‑range many‑body effects beyond the training window remain untested. Only monovalent 1:1 electrolytes are considered; multivalent salts, which are known to enhance strong coupling, are left for future work. The linear‑regression model, while efficient, may lack the expressive power of deep neural networks or graph‑based potentials for capturing more complex anisotropic correlations. Finally, explicit solvent effects (polarizability, specific ion adsorption) are ignored, which could be crucial for real experimental systems.

In summary, the authors demonstrate that a machine‑learned many‑body colloid‑only potential can faithfully reproduce primitive‑model forces and enable large‑scale simulations that were previously infeasible. Their results show that gas‑liquid spinodal instabilities appear only in the strong‑coupling regime of highly charged colloids, providing a clear computational benchmark for interpreting experimental observations of like‑charge attraction and highlighting the need for theories that incorporate strong electrostatic correlations rather than relying solely on Poisson‑Boltzmann mean‑field approximations.