I'm a machine learning researcher passionate about decoding complex systems—from brainwaves to biology. My work combines deep learning, signal processing, and simulation to understand and model temporal phenomena in neuroscience and the environment.

• Currently working at Beacon Biosignals, building models on EEG data for neurology.
• Previously researched wildfire dynamics in the Fire Management & Advanced Analytics Center, using landscape simulation, spatial statistics and deep learning.
• Interested in self-supervised learning, generative models, and representation learning for healthcare and biology.
• Skilled in Python, PyTorch, scikit-learn, and scientific computing.

Paper and Master's Thesis

In this paper, we propose a Deep Reinforcement Learning (DRL) framework to optimize the placement of firebreaks—structures that halt wildfire propagation—across large forested areas. Our approach offers a scalable and interpretable alternative to stochastic optimization or mixed-integer programming techniques, which become computationally intractable at scale.

• Formulated firebreak placement as a sequential decision-making problem.
• Trained a DQN agent on simulated wildfire data from Cell2Fire.
• Compared results to benchmark heuristic and optimization-based methods.

• Achieved competitive performance against mixed-integer optimization baselines.
• DRL agent learned spatial patterns of fire propagation and placed firebreaks in high-impact regions.
• The agent's improvement ranges between 1.59%–1.7% with respect to the heuristic, depending on the size of the instance, and 4.79%–6.81% when compared to a random solution.

• Reinforcement Learning (DQN)
• Simulation-based training environment (Cell2Fire)
• Forest grid encoding and episodic decision-making

Applied Soft Computing, Elsevier (2025)
DOI: 10.1016/j.asoc.2025.111234

Using Generative Models to Improve Wildfire Management

This project introduces a novel framework for improving forest firebreak placement using Graph Variational Autoencoders (VAEs) combined with two-stage stochastic optimization. Instead of randomly selecting simulation scenarios for optimization, we identify a compact and representative subset using latent space clustering—boosting both performance and generalization.

• Encoded 50,000 fire spread simulations (from Cell2Fire) into low-dimensional latent spaces using:
  • VGAE (Kipf & Welling, 2016)
  • GraphVAE (Simonovsky & Komodakis, 2018)
• Used K-Medoids clustering to extract representative graphs in latent space.
• Integrated with existing optimization model (Vilches et al., 2023) for strategic firebreak allocation.
• Validated improvements in wildfire mitigation over random sampling.

• 3.8% to 5.6% reduction in burned forest area across 1,000 evaluation simulations.
• Statistically significant performance gains (p < 0.05 for most experiments).
• Cluster-selected scenarios captured higher-magnitude fires, improving optimization robustness.

• Graph encoding with PyTorch Geometric
• Variational autoencoders for DAGs
• UMAP + K-Medoids for latent structure discovery
• Stochastic optimization via two-stage MILP

Latent Representation Meets Genetic Search

A novel research project combining latent space encoding with genetic algorithms for search and optimization. This hybrid approach explores how learned neural representations can inform and improve evolutionary strategies.

• Uses VAEs to compress complex data into search-ready representations.
• Mixes convenient properties of those representations to enhance genetic-based search.

U-Net + Grad-CAM for Wildfire Prediction

This project explores the interpretability of deep learning models trained to predict wildfire spread. A U-Net segmentation model is used to simulate wildfire front propagation, while Grad-CAM helps reveal which features (terrain, vegetation, etc.) drive extreme fire predictions.

• Spatially aware CNN architecture (U-Net).
• Grad-CAM integration for saliency and visual insights.
• Aimed at better understanding wildfire behavior through interpretable ML.

Efficient Open-Domain Question Answering

An implementation of Dense Passage Retrieval (DPR) fine-tuned using Low-Rank Adaptation (LoRA) for open-domain question answering.

• Based on Karpukhin et al.'s DPR (2020) and Hu et al.'s LoRA (2021).
• Fine-tuned on MSMARCO to boost top-K retrieval accuracy.
• Efficient adaptation of large models for real-world QA tasks.

• 📧 lucasmurrayh at gmail dot com