Coarse-grained models make large systems and long timescales accessible by representing groups of atoms with simplified interaction sites, but coarse graining necessarily removes atomistic information. Rather than assigning a single deterministic atomistic configuration to each coarse-grained structure, I develop autoregressive normalizing-flow decoders that sample ensembles of physically plausible atomistic configurations consistent with a given coarse-grained state. Using diphenylalanine (FF) as the model system, the models capture conformational diversity, preserve correlations between internal degrees of freedom, and produce structures that can be used directly in molecular dynamics.

A machine-learned structure is not treated as physically correct simply because it resembles the training set. Generated structures are constrained by molecular geometry and energetics and evaluated with molecular mechanics and statistical-mechanical reweighting, so their statistical weights remain physically interpretable — directly connecting modern generative modeling to the statistical-mechanical framework underlying molecular simulation. Manuscript in preparation: BoltzMap: Probabilistic Backmapping of Coarse-grained Peptides with Normalizing Flows (A. Kumar, S. Paschall, J. Monroe, 2026).

Backmapping is intrinsically a high-variance importance-sampling problem, and naive reweighting suffers from weight collapse. I study stabilization strategies — configurational-bias selection among candidate structures, restrained molecular-dynamics refinement, and careful diagnostics — that raise the effective sample size while keeping thermodynamic observables unbiased.

Aberrant protein aggregation underlies several neurodegenerative diseases, and it is a demanding computational problem: aggregation involves slow collective processes that are hard to capture with atomistic molecular dynamics, while the interactions that stabilize particular aggregates often need atomistic resolution to describe. Statistically rigorous transfer of information between coarse-grained and atomistic descriptions lets aggregation-prone conformations, transient intermediates, and sparsely populated states be studied at a combination of scales that neither representation achieves alone. The same methods extend to protein–nucleic-acid assemblies and biomolecular condensates.

Demonstrated by molecular simulation that a binary mixture of nanoparticles with simple isotropic interactions can self-assemble into ordered mesophases — lamellar, hexagonal, and double gyroid — previously associated with covalently connected block copolymers ("liquid crystals from spherical particles"). The key design is non-additive, frustrated attraction that sets up competition between mixing and demixing. Follow-up work mapped the full phase diagram, explained why the gyroid is harder to nucleate than lamellar and hexagonal phases, and followed the nucleation pathway from disordered liquid to gyroid.

Key publications: J. Phys. Chem. Lett. 2017 (ACS Most Read Physical Chemistry Articles of the Month); J. Phys. Chem. B 2018; J. Chem. Phys. 2019. Independent impact: the interaction model is referred to by other researchers as the Kumar–Molinero model and, since 2018, has been independently used, reproduced, benchmarked, adapted, or extended by at least five research groups in the United States, Argentina, and the Republic of Korea.

Extending the KM framework with tetrahedral interactions (a network former plus a structure-directing agent) produced a coarse-grained model of zeolite formation in which six distinct zeolite polymorphs crystallize, five of them through metastable mesophase intermediates. These first-author studies showed that mesophases can precede and direct zeolite nucleation, provided a mechanistic framework for polymorph selection in porous materials, and revealed a two-step to one-step transition in zeolite nucleation through a metastable gyroid mesophase. The zeolite work was supported by the ACS Petroleum Research Fund, and its results provided the preliminary data for a subsequent U.S. Department of Energy project on zeolite formation in the Molinero group.

Key publications: J. Am. Chem. Soc. 2018; J. Phys. Chem. Lett. 2018 (ACS PRF 25th-Anniversary Collection, 2025); J. Phys. Chem. C 2019. Independent impact: the coarse-grained zeolite model was adopted by researchers at Utrecht University as the basis of an inverse-design workflow that reproduced known frameworks (Z1, SGT), optimized others (SOD, CFI), and reported an additional framework, Z5 (ACS Nano 2025).

Simulations of the intrinsically disordered protein FUS showed that sequence determines the switch in its fibril-forming regions and variants, revealing distinct fibril-core morphologies, sequence-dependent stability differences, and the impact of phosphorylation on disordered proteins. In parallel, coarse-grained (SOP-IDP) force fields for intrinsically disordered proteins were optimized and their conformational ensembles validated against experimental SAXS profiles, improving the structural characterization of IDPs.

Key publications: J. Phys. Chem. Lett. 2021 (first author); Biopolymers 2024; Protein Science 2025.

Coarse-grained HPS simulations (HOOMD-blue) of HP1–DNA condensates elucidating the thermodynamic stability and diffusion dynamics that underlie chromatin compaction. Presented at the Biophysical Society Annual Meeting (Philadelphia, 2024).