Title: Machine learning glasses and their enhanced sampling
Abstract:
It remains a difficult task to predict from first principles the physical properties of disordered materials that are formed out of equilibrium, because the lack of an apparent order and the loss of ergodicity severely limit the use of statistical mechanics and computational tools. As in many fields, various machine learning strategies are being developed to attack these open physics problems with new forces. In this talk, I will explain where and how machine learning techniques can help answer some of these important questions to understand the physics of the glass transition and the properties of amorphous solids.
References:
G. Jung, G. Biroli, and L. Berthier, Mach. Learn.: Sci. Technol. 5, 035053 (2024).
G. Jung, et al., Nat. Rev. Phys. 7, 91 (2025)
G. Jung, G. Biroli, and L. Berthier, Phys. Rev. Lett. 130, 238202 (2023).
Title: Machine Learning Approaches to Energy Landscape Exploration in Disordered Materials
Abstract:
The energy landscape paradigm provides a powerful framework for understanding the structure, dynamics, and properties of disordered systems such as glasses. However, navigating these complex, high-dimensional landscapes remains computationally challenging. In this talk, we will discuss our recent advances in leveraging machine learning and differentiable simulations to explore and exploit energy landscapes of disordered materials. First, we will introduce StriderNet, a reinforcement learning framework designed for efficient exploration of potential energy landscapes. StriderNet enables rapid identification of local minima and transition states, accelerating the discovery of stable and metastable configurations in glasses and other disordered systems. Second, we will demonstrate our approaches for visualizing high-dimensional energy landscapes through dimensionality reduction and interpretable machine learning techniques. These visualization methods reveal hidden patterns in configurational space and provide intuitive understanding of energy barriers, basins, and pathways that govern material behaviour. Finally, we will present our work on differentiable molecular dynamics simulations that enable direct estimation of transport properties such as viscosity. By making the simulation process end-to-end differentiable, we can compute gradients of macroscopic properties with respect to structural features, opening new avenues for inverse design and property prediction in glassy materials. Together, these advances demonstrate how modern computational techniques are transforming our ability to understand and engineer disordered materials through their energy landscapes.
Title: Beyond Propagation of Chaos: A Stochastic Algorithm for Mean Field Optimization
Abstract:
Gradient flow in the 2-Wasserstein space is widely used to optimize functionals over probability distributions and is typically implemented using an interacting particle system with n particles. Analyzing these algorithms requires showing (a) that the finite-particle system converges and/or (b) that the resultant empirical distribution of the particles closely approximates the optimal distribution (i.e., propagation of chaos). However, establishing efficient sufficient conditions can be challenging, as the finite particle system may produce heavily dependent random variables. In this work, we study the virtual particle stochastic approximation, originally introduced for Stein Variational Gradient Descent. This method can be viewed as a form of stochastic gradient descent in the Wasserstein space and can be implemented efficiently. In popular settings, we demonstrate that our algorithm's output converges to the optimal distribution under conditions similar to those for the infinite particle limit, and it produces i.i.d. samples without the need to explicitly establish propagation of chaos bounds.
Bruno Loureiro
Title: Scaling Laws and Spectra of Shallow Neural Networks in the Feature-Learning Regime
Abstract:
Neural scaling laws underlie many of the recent advances in deep learning, yet their theoretical understanding remains largely confined to linear models. In this work, we present a systematic analysis of scaling laws for quadratic and diagonal neural networks in the feature learning regime. Leveraging connections with matrix compressed sensing and LASSO, we derive a detailed phase diagram for the scaling exponents of the excess risk as a function of sample complexity and weight decay. This analysis uncovers crossovers between distinct scaling regimes and plateau behaviours, mirroring phenomena widely reported in the empirical neural scaling literature. Furthermore, we establish a precise link between these regimes and the spectral properties of the trained network weights, which we characterize in detail. Consequently, we provide a theoretical validation of recent empirical observations connecting the emergence of power-law tails in the weight spectrum with network generalization performance, yielding an interpretation from first principles.
References:
Hugo Cui et al J. Stat. Mech. (2022) 114004.
Leonardo Defilippis and Bruno Loureiro and Theodor Misiakiewicz, arxiv.org/abs/2405.15699 (2024).
Leonardo Defilippis et al. , arxiv.org/abs/2509.24882 (2025).
Title: Sampling in the Collective Variable Space
Abstract:
We consider the problem of sampling a high-dimensional multimodal target probability measure. We assume that a good proposal kernel to move only a subset of the degrees of freedoms (also known as collective variables) is known as a priori. This proposal kernel is typically built in practice using normalizing flows. We show how to extend the move from the collective variable space to the full space and how to implement an accept-reject step in order to get a reversible chain with respect to a target probability measure. The accept-reject step does not require to know the marginal of the original measure in the collective variable (namely to know the free energy). The obtained algorithm admits several variants, some of them being very close to methods which have been proposed previously in the literature. We show how the obtained acceptance ratio can be expressed in terms of the work which appears in the Jarzynski-Crooks equality, at least for some variants. Numerical illustrations demonstrate the efficiency of the approach on various simple test cases and allow us to compare the variants of the algorithm.
References:
This is a joint work with Marylou Gabrié, Christoph Schoenle, and Gabriel Stoltz.
Title: Building Good Representations of Local Environment with Rotation-Equivariant Neural Networks
Abstract:
In this talk I will introduce the key concepts about group-equivariant neural networks [Thomas 2018], in particular for the case of the 3D rotation+translation group, SE(3), and try to open perspectives for future research. I will first define and compare the notions of group-equivariance and group-invariance, motivating our use of roto-translational equivariant features (representations). Skipping most of the technical group theory/Spherical Harmonics part, starting mostly from intuitions, I will show how one can in practice define a steerable network (i.e. equivariant by design), while keeping expressivity, despite the strong constraint of equivariance. I will then mention our main results in terms of generalization to temperatures different from the training set [Pezzicoli 2024], and how these generalization results open perspectives for a learned tensorial order parameter.
If time allows, I will mention the very expressive MACE architecture [Batatia 2022] and its successes.
References:
Tensor field networks: Rotation- and translation-equivariant neural networks for 3D point clouds, Google Research(2018)
Rotation-equivariant graph neural networks for learning glassy liquids representations, SciPost 2024
MACE: Higher order equivariant message passing neural networks for fast and accurate force fields, NeurIPS 2022
Title: Predicting Force Chains in Jammed Solids Using Machine Learning
Abstract:
Force chains are quasi-linear self-organised structures that carry large stresses and are ubiquitous in jammed amorphous materials like granular solids, foams, or even cell assemblies. Predicting where they will form upon deformation remains an open question. Here we demonstrate that graph neural networks (GNN) can accurately predict the location of force chains in both frictionless and frictional materials from the undeformed structure, without any additional information. The GNN prediction accuracy also proves to be robust to changes in packing fraction, mixture composition, amount of deformation, friction coefficient, system size, and the form of the interaction potential. Our results and methodology will be of interest for granular matter and disordered systems, e.g., in cases where direct force chain visualization or force measurements are impossible.
Title: Ambient and High-Pressure B₂O₃ Glass
Abstract:
Despite their ubiquity, inorganic glasses remain outliers to the structure-property paradigm that unites molecular and crystalline sciences. The lack of model-free experimental and reliable computational access to interatomic structure in glasses is the impediment for this situation. MD simulations with empirical force fields for such extended network systems often over stabilize cohesive forces; with very little validation beyond matching peak positions against experimental X-ray/neutron scattering patterns, their predictions of interatomic structure do not match up to the progress made with MD, say, in its description of molecular liquids. On the other hand, ab initio MD simulations (AIMD) (based on, say, DFT as the ground truth) are far behind in the timescales required for structural relaxation in supercooled liquids. Recent developments in machine-learned force fields have been able to address a few of these issues, providing a neural network representation of DFT-derived atomic forces and total energy of a configuration1. The current work demonstrates the use of the Deep Potential for a classic extended network glass, B2O3. Our work2 shows (i) the formation of denser glasses with decreasing quench rate, (ii) quench rates larger than 1011 K/s leads to the formation of artifactual four-membered rings in the structure (iii) the formation of tetrahedral BO4 units at high pressures. We shall also cover the issue of the fraction of boron atoms present in hexagonal boroxol rings3-5, an entity that has remained elusive to MD simulations so far.
References:
V. L. Deringer, N. Bernstein, A. P. Bartók, M. J. Cliffe, R. N. Kerber, L. E. Marbella, C. P. Grey, S. R. Elliott, and G. Csányi, Realistic Atomistic Structure of Amorphous Silicon from Machine-Learning-Driven Molecular Dynamics, J. Phys. Chem. Lett. 9, 2879 (2018).
D. Meher, N.V.S. Avula, and S. Balasubramanian, Slowly quenched, high pressure glassy B2O3 at DFT accuracy, J. Chem. Phys., 162, 044503 (2025).
P. Umari and A. Pasquarello, Fraction of Boroxol Rings in Vitreous Boron Oxide from a First-Principles Analysis of Raman and NMR Spectra, Phys. Rev. Lett. 95, 137401 (2005).
G. Ferlat, “Rings in network glasses: The B2O3 case,” in Molecular Dynamics Simulations of Disordered Materials: From Network Glasses to Phase-Change Memory Alloys, edited by C. Massobrio, J. Du, M. Bernasconi, and P. S. Salmon (Springer International Publishing, Cham, 2015), pp. 367–414.
D. Meher, N.V.S. Avula, and S. Balasubramanian, Intermediate Range Order and Boroxol Fraction in Melt-Quenched B2O3 Glass Modeled at DFT Accuracy (Communicated) (2025).
Misaki Ozawa
Title: Wavelet Conditional Renormalization Group and Its Applications in Physics
Abstract:
We develop a wavelet conditional renormalization group (WCRG) framework that constructs a generative model for physical data with high efficiency [1]. By operating in a hierarchical, scale-by-scale manner, the method captures the underlying physical mechanisms embedded in complex datasets. We demonstrate its success on the phi4 model at thermal equilibrium, as well as on non-equilibrium astrophysical data. Furthermore, our approach enables the estimation of the Shannon entropy for non-equilibrium systems where the underlying Hamiltonian is unknown a priori, providing a new route to characterize such systems from data alone [2].
References:
Marchand, Ozawa, Biroli, Mallat, Physical Review X 13, 041038 (2023).
Martiniani, Chaikin, and Levine, Physical Review X 9, 011031 (2019).
Tarak Karmakar
Title: Enhanced Sampling with Machine Learning Interatomic Potentials for Rare Event Simulations
Abstract:
Modelling ‘rare events’ – such as chemical reactions, diffusion, and conformational transitions – remains one of the major challenges in molecular simulations due to the presence of kinetic bottlenecks having high free-energy barriers and the standard brute-force simulation’s inability to access long timescales. Traditional ab initio molecular dynamics (AIMD) simulations, while accurate, are computationally prohibitive for exploring such processes extensively with affordable computational resources. In our work, we combine machine learning interatomic potentials (MLIPs) with enhanced sampling (ES) techniques to achieve quantum-level accuracy at computational efficiency comparable to classical simulations. MLIPs, developed using the Deep Potential Molecular Dynamics (DeepMD) framework and trained on high-quality density functional theory (DFT) data, accurately capture reactive potential energy surfaces for complex materials and catalytic systems. By integrating these MLIPs with ES simulations such as Well-tempered Metadynamics (WTMetaD) and On-the-fly Probability-based Enhanced Sampling (OPES), we efficiently sample rare events, including surface catalysis, nanocluster dynamics, and gas transport in confined materials that were previously inaccessible to direct AIMD simulations. The results highlight how ML-driven enhanced sampling can bridge the gap between accuracy and efficiency, offering a powerful route for predictive modelling of complex chemical and materials processes.
References:
Anmol, Sirohi A. K., Neha, Jayadeva, Kumar S., Karmakar T. J. Chem. Theory Comput. 2025, 21, 12, 6113–6120.
Tiwari V., Gupta D., Karmakar T. ChemRxiv, 2025.
Tiwari V., Karmakar T. Nano Lett. 2025, 25, 14, 5940–5946.
Karmakar A., Dhananjay, Karmakar T. Nano Lett. 2025, 25, 25, 10187–10192.
Title: Buckets Instead of Umbrellas for Enhanced Sampling and Free Energy Calculations
Abstract:
Umbrella sampling has been a workhorse for free energy calculations in molecular simulations for several decades. In conventional umbrella sampling, restraining bias potentials are strategically applied along one or several collective variables. Major drawbacks associated with this method are the requirement of a large number of bias windows and the poor sampling of the transverse coordinates. In my presentation, I will introduce an alternate formalism that departs from the traditional umbrella sampling to mitigate these issues, where we replace umbrella-type restraining bias potentials with bucket-type wall potentials. This modification permits one to formulate an efficient computational strategy leveraging wall potentials and meta-dynamics sampling. This new method, called “bucket sampling”, can significantly reduce the computational cost of obtaining converged high-dimensional free energy surfaces. Extensions of the proposed method with temperature acceleration and replica exchange solute tempering will be also demonstrated.
References:
Javed, R.; Kapakayala, A. B.; Nair, N. N. J. Chem. Theory Comput. 2024, 20 (19),8450-8460.
Jagannath Mondal
Title: Enhancing Sampling of Rough Landscapes via Transformer and Diffusion Models
Abstract:
Rugged free-energy landscapes hinder efficient exploration by conventional MD. I’ll present a set of complementary AI routes. First, denoising diffusion probabilistic models trained on short MD segments generate atomistically plausible conformations and reach sparsely populated states, improving coverage across folded and disordered proteins while clarifying limits in rare basins. Second, a decoder-only transformer treats discretized trajectories as “language,” forecasting kinetically consistent state sequences and extending sampling along the time axis. Together, diffusion broadens configurational reach while transformers advance temporal progression, offering a practical, validated toolkit for rough landscapes. We will be describing promising directions of using Generative adversarial neural network (GAN) to enhance sampling in a time-sequenced manner.
References:
P. Bera and J. Mondal Chemical Science 16, 8735 (2025).
P. Bera and J. Mondal J.Chem.Phys. 163,114110 (2025).
Title: Engineering Polymer-Surface Adhesion Using Molecular Dynamics and Machine Learning
Abstract:
Designing of functional polymers with tailored applications requires an understanding of the interactions with surfaces that find applications in chemical, biological and industrial processes. The effects of polymer (or protein) sequence and the surface composition are expected to be essential in determining the adsorption of bio-polymers on the surface. Examples vary from the sequence-dependent aggregation of proteins on the membrane to designing adhesive polymeric materials for industrial applications. Integrating molecular dynamics (MD) simulations and machine learning (ML) approaches can act as effective tools in understanding of sequence design principles. In this work, we compute and predict the adhesive free energies of generic, coarse-grained polymers to generic functionalized surfaces having variable sequences, composition, spatial patterns and polymer-surface interaction strengths. Enhanced molecular dynamics simulations (approximately 0.85 million simulations) are employed to generate a unique synthetic dataset of 8464 polymer-surface binding free energies (BFEs). The computed BFEs account for the effects of polymer sequences, polymer/surface composition and variable polymer-surface interaction strengths. This dataset is used for training the CNN-GRU attention-based hybrid model that accurately predicts the BFEs as well as complete free energy profiles of the polymers binding to surfaces with unknown sequences and composition. This approach is also extended towards inverse design of polymers and surfaces based on desired adhesion properties. Our work provides an unified approach using MD simulations with ML to accelerate the design of functional polymers for various interfacial applications.
References:
Ozboyaci M, Kokh DB, Corni S, Wade RC. Q. Rev. Biophys. 2016;49:e4.
Brancolini, G.; Corazza, A.; Vuano, M.; Fogolari, F.; Mimmi, M. C.; Bellotti, V.; Stoppini, M.; Corni S.; Esposito, G. ACS Nano 9, 2600–2613.
Rouse, I.; Lobaskin, V. Faraday Discuss., 2023, 244, 306–335.
Shi, J.; Quevillon, M. J.; Valenca, P. H. A.; Whitmer , J. K. ACS Appl. Mater. Interfaces 2022, 14, 37161−37169.
Panigrahy, S.; James, A; Nayar, D. AI for Accelerated Materials Design- NeurIPS 2025, 2025.
Tony Bonnaire
Title: The Role of the Time-Dependent Hessian in High-Dimensional Optimization
Abstract:
How does gradient descent (GD) manage to find good solutions in rough, high-dimensional landscapes? We address this question through the phase retrieval problem, a simple but representative non-convex model. By tracking the Hessian spectrum during training, we uncover a dynamical transition that governs how GD escapes rugged regions of the loss landscape. At sufficiently large signal-to-noise ratio (SNR), an informative negative direction in the Hessian of initial states guides the dynamics towards a good solution. However, if the SNR is not large enough, this direction is lost while descending through a BBP transition and the system becomes trapped into high-energy minima. Interestingly, for finite-size systems, this window of negative curvature is enough to recover the signal below the asymptotic SNR threshold, showing how initialization and early-time dynamics crucially shape learning in rough landscapes.
References:
Bonnaire et al., J. Stat. Mech. (2025) 083401.
Sarao Manelli et al., NeurIPS (2020).
Title: Neural Large Deviations Generator
Abstract:
Barrier crossings, current fluctuations, and entropy bursts are typically associated with rare trajectories in stochastic systems. Rare trajectories are naturally described by the biased ensembles of large-deviation theory [1]. Sampling such ensembles is computationally expensive, as it requires calculation of the dominant eigenvalues of tilted generators [2] or generating exponentially many trajectories. We introduce a neural generative framework that learns the equilibrium trajectory distribution and then constructs biased large-deviation ensembles by score tilting and flow matching in the trajectory space. Starting from a diffusion-model [3], the learned score field s(Y) = ∇_Y log P(Y) represents the gradient of the log-probability density of full trajectories Y={x_t} t∈[0,T] under the original dynamics. Rare trajectories associated with a time-averaged observable A_T [Y]=1/T∫^T_0 a(x_t)dt
are generated by modifying the score as
sbY=sY+b∇Y ATY,
This corresponds to the first-order Doob transform of the underlying dynamics. We then extend this to finite bias b by training a flow-matching Schrodinger bridge ¨ [4] that continuously transports P[Y] into the biased ensemble P_b[Y] ∝ e^b A_T [Y]P[Y].
In prototype systems of one and two-dimensional dimensions, the method reproduces the expected shift of observables averaged over time and produces accurate scaled cumulant generating functions λ(b) across a finite bias range. Our approach thus provides an amortized pathway to learn and generate large-deviation trajectories from equilibrium data, bridging tilted statistical physics operators and generative modelling frameworks.
References:
H. Touchette, The large deviation approach to statistical mechanics, Physics Reports, 478(1–3),1–69 (2009).
R. Chetrite and H. Touchette, Nonequilibrium microcanonical and canonical ensembles and their equivalence, Physical Review Letters, 111, 120601 (2013).
L. Yang, Z. Zhang, Y. Song, et al., Diffusion Models: A Comprehensive Survey of Methods and Applications, ACM Computing Surveys (2023).
De Bortoli, Valentin and Albergo, Michael S. and Vanden-Eijnden, Eric, Flow Matching for Generative Modelling
Title: Stochastic Interpolants, Diffusion Models, and the Exact Renormalization Group
Abstract:
-
Title: Rugged Yet Accessible: The Topography of a Tractable Fitness Landscape
Abstract:
The adaptation of an organism to environmental stress often involves a loss of fitness in the stress-free environment. Upon removal of the stress, this loss can be mitigated by compensatory evolution through mutations that are not involved in the stress resistance mechanism, rendering the organism well-adapted to environments both with and without the stress. In this talk I will present an evolutionary model of a fitness landscape that reproduces such adaptive behaviour. The landscape is rugged, yet its peaks are accessible. In addition, the adaptive processes on this landscape have parallels with the evolution of plasticity in amorphous materials.
References:
S.G. Das, S.O.L Direito, B. Waclaw, R.J. Allen, and J. Krug, "Predictable properties of fitness landscapes induced by adaptational tradeoffs," Elife, 9:e55155, 2020.
S. G. Das, M. Mungan and J. Krug, "Epistasis-mediated compensatory evolution in a landscape with adaptational tradeoffs,” Proc. Natl. Acad. Sci. USA 122 (2025) e2422520122
Title: Discovery of Unconventional and Novel Materials Using Machine Learning
Abstract:
Machine learning (ML) approaches have demonstrated remarkable potential for discovering new materials across diverse domains, including catalysis, energy storage, and biomolecular design. However, these methods are often criticized for being primarily interpolative, enhancing material performance within known boundaries rather than uncovering truly novel systems. In this talk, I will illustrate, using the example of β-sheet formation in peptides, how ML methods can successfully result in identification new materials with desired properties which are not only unconventional but also cannot be easily discovered using traditional domain-specific physics models. I will further highlight how active learning strategies can be leveraged to target regions where ML predictions and conventional models disagree, enabling the discovery of genuinely novel materials. Finally, I will discuss how integrating transformer-based large language models (LLMs) with traditional, information-rich cheminformatics descriptors can significantly improve polymer property prediction, offering a pathway toward more interpretable and generalizable material design frameworks.
References:
Talluri, Y. Nissi, et al. Science Advances 11.24 (2025): eadt9466.
Parambil, Vijith, et al. Materials Today Communications (2025): 113375.
Suman Chakrabarty
Title: Enhanced Sampling and Estimation of Rare-Event Kinetics in Molecular Simulation: Some Ideas and Benchmarks
Abstract:
Understanding the molecular mechanism of any complex biophysical or chemical processes requires tracking of the dynamics on the underlying (free) energy landscape. For all practical purposes, this requires projection from a higher-dimensional landscape (3N for N particle system) onto a few “collective variables” or order parameters or reaction coordinates. The accuracy of the computed free energy landscape and barrier often strongly depends on this choice of collective variables to project on. Moreover, estimation of rare event kinetics (rate constant) from the underlying free energy landscape suffers from the lack of any exact rate theory. In this talk, we shall discuss the “weighted ensemble” (WE) class of path sampling method for directly estimating rate constants of rare events without any biasing forces or elevated temperature and without pre-computing the underlying free energy barrier. We shall introduce a few variants of WE-based methods developed in our group.
References:
WeTICA: A directed search weighted ensemble based enhanced sampling method to estimate rare event kinetics in a reduced dimensional space, S. Mitra, R. Biswas, and S. Chakrabarty, J. Chem. Phys. 162, 034106 (2025).
IceCoder: Identification of Ice Phases in Molecular Simulation Using Variational Autoencoder, D. Maity, and S. Chakrabarty, J. Chem. Theory Comput. 21, 1916 (2025).
Mapping conformational landscape in protein folding: Benchmarking dimensionality reduction and clustering techniques on the Trp-Cage mini-protein, S. Bhattacharya, and S. Chakrabarty, Biophysical Chemistry 319, 107389 (2025).
Title: Generative AI for Exploring and Enhancing Sampling of Chemical Space
Abstract:
-