Machine Learned Landscape Flattening (mllf)
Brief introduction
Machine Learned Landscape Flattening (mllf) provides tools and documentation for applying adaptive biasing and multisite λ-dynamics to accelerate alchemical free-energy calculations. The project includes a contextual bandit (CB) framework using graph neural networks to predict optimal bias coefficients, complete training workflows, and utilities for working with simulation outputs.
What you’ll find in these docs
Background: conceptual and mathematical background for multisite λ-dynamics and Adaptive Landscape Flattening (ALF)
Installation: how to install the package and dependencies
File Handling: parsers and writers for RTF, PDB, bias coefficients, and output files
DeepSet Pretraining: learned physical representations from atomic structure
CB Behavior Cloning: transfer learning from expert bias coefficients
Contextual Bandit Setup: graph neural network architecture for bias prediction
Workflow System: complete pipeline from combination generation to training
Examples: runnable training workflows with SLURM integration
References: bibliography for cited literature
API: generated API reference for the
mllfpackage
Quick start (recommended)
Read the Background section to understand the modeling approach
Follow Installation to set up the environment
Review File Handling to understand input/output formats
Review Workflow System for the complete pipeline overview
(Optional) Read DeepSet Pretraining and CB Behavior Cloning for advanced features
Check Examples for running the training workflow
Inspect the API reference for integration details