Materials informatics, in depth.
GNN-based property prediction, Bayesian active learning, transfer learning from DFT to experiment, and practical ML for battery, polymer, and alloy R&D teams.
Active Learning for Synthesis Planning: Fewer Experiments, Faster Convergence
How Bayesian active learning acquisition functions decide which experiment to run next, and why this matters more than the size of your training set.
Read article
The Small-Data Problem in Materials ML: Why 200 Data Points Can Be Enough
Materials datasets are tiny by ML standards. Here is what actually works when you have 50 to 500 experimental measurements.
Composition Optimization for Structural Alloys: A Worked Example
A walkthrough of how Matforgelab maps a fatigue-life and yield-strength specification to ranked alloy candidates.
Predicting Polymer Glass Transition Temperature from Monomer Composition
Tg prediction from composition using graph-based representations. What the model learns and where it struggles.
Closing the Experimental Loop: How Measured Data Sharpens Future Predictions
What happens when you feed characterization results back into the model: how posterior updates work in practice.
Graph Neural Networks for Materials: Why Composition Alone Is Not Enough
How GNNs encode crystal and molecular graphs, and what this buys you over descriptor-based methods.
Transfer Learning from DFT to Experiment: Closing the Simulation Gap
Pre-training on DFT-computed properties and fine-tuning on experimental measurements: why it works and where it breaks.
High-Throughput Electrolyte Screening: The Case for Prediction Before Synthesis
Electrolyte formulation space for solid-state batteries is vast. How prediction-driven screening changes the campaign economics.
Predicting Cathode Cycle-Life from Composition: What the Model Sees
Cycle-life is one of the hardest cathode properties to predict. Here is what ML can and cannot tell you.
Why R&D Teams Spend 70% of Their Synthesis Budget on Candidates That Will Not Work
The formulation screening bottleneck is not a resources problem. It is an information problem.