01
Rare-event sampling
Adaptive simulation strategies for efficiently discovering infrequent molecular transitions.
Research overview
Rare events, pathways, and learned representations
I use representation learning and statistical mechanics to construct data-driven collective variables and to map complex free-energy landscapes, and I build open-source scientific software that makes these methods accessible for molecular dynamics and materials-simulation workflows.
Research
A synthesis of the research profile
My research develops and applies machine-learning-guided methods for sampling and characterising rare events in molecular systems, with emphasis on adaptive and weighted-ensemble sampling strategies for transition-path discovery and pathway-resolved kinetics.
02
Transition pathways & kinetics
Pathway discovery, pathway-resolved analysis, weighted-ensemble simulations, fluxes, MFPTs, and kinetic observables.
03
Machine-learned molecular representations
Autoencoders, variational autoencoders, learned collective variables, dimensionality reduction, and data-driven descriptors.
04
Molecular & materials simulation
Applications spanning proteins, molecular recognition, phase transitions, molecular materials, solvation, and related problems.
Featured research
Methodological contributions
PathGennie
Direction-guided adaptive sampling for rare-event transition pathways
- Problem
- Rare molecular transitions — folding events, phase changes, chemical reorganizations — occur on timescales far beyond what direct molecular dynamics can reach, making transition pathways difficult to sample efficiently.
- Approach
- PathGennie uses ultrashort monitored trajectories to steer adaptive sampling along a direction that biases exploration toward the transition, generating rare-event pathways rapidly without long unbiased runs.
- Contribution
- Provides a practical, direction-guided adaptive-sampling scheme for generating transition pathways at substantially reduced computational cost, released as an open-source Python framework.
IceCoder
Variational-autoencoder representation learning for ice-phase identification
- Problem
- Distinguishing ice polymorphs and liquid environments from local molecular geometry is difficult with hand-crafted order parameters, which are often specific to a single phase or system.
- Approach
- IceCoder learns a compact latent representation of local molecular environments with a variational autoencoder, trained without phase labels, so that structurally distinct environments separate naturally in latent space.
- Contribution
- Delivers an unsupervised classifier for ice phases and liquid environments that generalizes across simulation conditions, packaged as an open-source PyTorch framework.
Pathway-resolved kinetics
Neural-network-guided weighted-ensemble simulations for channel-specific rates
- Problem
- When a rare event proceeds through multiple competing pathways, standard weighted-ensemble kinetics estimates the overall rate but obscures how much flux flows through each individual channel.
- Approach
- Candidate transition routes are grouped and refined into smooth reference pathways described by a neural-network path collective variable (PathCV), and independent weighted-ensemble (WE) simulations are then run along each channel.
- Contribution
- Enables quantitative, pathway-resolved kinetics — channel-specific fluxes and mean first-passage times — from weighted-ensemble simulations rather than a single lumped rate.
TRAILS-MD
Lineage-aware adaptive molecular-dynamics framework
- Problem
- Adaptive sampling campaigns spawn and prune many trajectory segments over time, and tracking how each segment relates to its ancestry is essential for correct reweighting and analysis.
- Approach
- TRAILS-MD runs many short, segmented parallel walkers and records which earlier segment produced each new one, so continuous transition paths can be reconstructed from many short pieces regardless of the underlying MD engine.
- Contribution
- Turns rapid, broad exploration into connected, interpretable transition pathways rather than a scatter of disconnected samples, released as reusable adaptive-sampling infrastructure.