I’m an Applied Physics PhD candidate at Rice University developing scientific software and predictive models for complex experimental data. My work spans microscopy, computational physics, biomedical data analysis, and high-performance computing.

I build validated, reusable workflows—from multidimensional imaging and physics-based simulations to cancer drug-response modeling and optimized native implementations.

• Cancer Drug Response Prediction A GDSC2 study of 91,229 analyzed drug–cell responses, asking how well gene expression and molecular fingerprints predict assay response in unseen cell lines for drugs represented in training. On 120 held-out cell lines, validation-selected per-drug Ridge reduced test RMSE 18.1% versus the per-drug training-mean baseline; a three-seed PyTorch MLP did not outperform per-drug Ridge. The repository includes cell-line-disjoint evaluation, 2,000 paired cell-line bootstrap resamples, drug-level error analysis, and data provenance. A later exploratory molecular GNN extension tested drug graphs on the same split and did not improve on Ridge.

• 4Denoise An open-source Python package for processing and visualizing 4D-STEM data. It supports preprocessing, Bragg-peak analysis, crystallographic tilt inference, 3D topography reconstruction, tilt-corrected strain mapping, and interactive visualization. The package supports my first-author Science Advances study, “Strain Mapping of Three-Dimensionally Structured Two-Dimensional Materials”.

• Generalized C-MEP A reusable Python workflow for correlative multislice electron ptychography, including physically calibrated 3D volumes, registration of complementary datasets, relative overlap scoring, subvoxel atom localization, and simulation-based validation with ASE and abTEM.

• CUDA C++ Performance Engineering (in active development) A correctness-first portfolio for scientific denoising with C++17, OpenMP, CUDA, and Python bindings validated against reference outputs. Benchmarks separate kernel performance from transfer-inclusive costs and document retained and rejected optimizations.

• Reinforcement Learning Control of a Double Pendulum A continuous-control project built from symbolically derived nonlinear equations of motion, custom Gym environments, and Stable-Baselines3 DDPG agents trained to control a double pendulum on a cart.

I use Codex and GPT coding assistants to turn research goals into bounded tasks, review implementations, build tests, and improve documentation. I independently validate numerical results and remain responsible for scientific assumptions, software decisions, and reproducibility.

• Scientific software, materials informatics, and biomedical data analysis
• Predictive modeling with rigorous validation
• AI-assisted software development
• C++ and CUDA performance engineering
• Computer vision and multidimensional image processing
• Correlative 3D reconstruction and inverse problems
• Physics-based machine learning
• Scientific visualization and interactive tools

When I’m not working with microscopy data or debugging code, I:

• 💪 Arm wrestle
• 🪂 Skydive
• 🎹 Play piano and ukulele
• 🍽️ Enjoy exploring new restaurants

• LinkedIn: Adan J. Mireles
• Email: mireles@rice.edu