Computational Physics | Machine Learning | Quantum Computing

I build physics simulation systems, ML pipelines, and scientific applications - work that sits between computational science and software engineering.

🚀 Live Demo

Multi-source academic search aggregator with GPT-powered analysis. Combines OpenAlex, Semantic Scholar, and arXiv (opt-in) into a single search interface with intelligent deduplication, citation enrichment, AI-generated research summaries and a RAG-based query pipeline. Generates both quick landscape overviews and in-depth 1-page syntheses to accelerate literature review workflows.

AWS Architecture: Next.js 14 static site hosted on AWS Amplify with CI/CD from GitHub. Backend uses AWS Lambda (Python) behind API Gateway for multi-source aggregation, with DynamoDB caching (7-day search cache, 24-hour deep overview cache). OpenAI API integration (GPT-4o-mini) for intelligent summarization.

Tech: TypeScript, Next.js 14, Tailwind CSS, Python, AWS Lambda, API Gateway, DynamoDB, AWS Amplify, OpenAI API

🚀 Demo

Discovers governing PDEs from spatial-temporal data using Physics-Informed Neural Networks and sparse regression. Upload experimental data and it identifies a candidate equation via PINN training and sparse regression over a library of 12 candidate terms (linear, nonlinear, higher derivatives). Single-service demo (FastAPI + SQLite) tested up to 10,000+ points; includes sample datasets (heat diffusion, Burgers, KdV equations).

Automates equation discovery from experimental physics data—turning workflows that are often manual and iterative into a reproducible pipeline.

Tech: Python, FastAPI, PyTorch, NumPy/SciPy, SQLite

Containerised geospatial solver demo exploring quantum-classical hybrid patterns. Implements a demonstration-scale HHL quantum linear solver (via Qiskit) with automatic fallback to classical sparse solvers; full job pipeline with web submission and PostGIS integration.

Tech: Python, Qiskit, FastAPI, PostgreSQL/PostGIS, Celery, Redis, Docker

PyQt6 desktop application for visualizing Earth's magnetic field using spherical harmonic decomposition. Processes sensor data in real-time, generates 3D field vector visualizations, and exports analysis reports. Built to explore geomagnetic field modeling with hardware integration via PySerial.

Tech: Python, PyQt6, NumPy, SciPy, Plotly, PySerial

Unsupervised anomaly detection framework that quantifies universality class proximity directly from finite-size simulation data. An Isolation Forest trained on Edwards-Wilkinson and KPZ surfaces identifies unknown dynamics (MBE, VLDS, Quenched-KPZ) as anomalous with 100% detection at L=128–512. Key contribution: a continuous universality distance D_ML(κ) that characterizes crossover behavior with 2× better signal-to-noise than traditional scaling exponent fitting.

Provides an operational, data-driven diagnostic for finite-size data where traditional power-law fitting is unreliable—addresses a practical problem in non-equilibrium statistical mechanics.

Key Results: Crossover scale κ_c = 0.76 ± 0.05, R² = 0.96. Feature ablation shows gradient/temporal statistics achieve 100% detection while scaling exponents (α, β) achieve only 79%.

Tech: Python, Scikit-learn (Isolation Forest), NumPy, SciPy, Numba (JIT-compiled simulations)

Applies the Wasserstein distance framework from my universality classification research to financial regime detection. An attempt to translate the mathematical machinery (optimal transport, observable maps, anomaly detection) from physics to finance. Computes rolling statistical features, measures distribution divergence, flags regime changes. Whether market regimes share the mathematical structure of universality classes remains an open question—this explores the translation.

Tech: Python, POT (Optimal Transport), Scikit-learn, NumPy, yfinance

Exploratory numerical study of two coupled surface growth interfaces (modified KPZ). Implemented a Numba-accelerated simulator plus analysis/visualization tools (cross-correlation, scaling diagnostics) and example parameter sweeps across coupling strengths and coupling symmetry.

Addresses a gap in non-equilibrium statistical mechanics research—explores coupling effects that could apply to thin film deposition, biological growth, and interface dynamics in materials science.

Tech: Python, NumPy, SciPy, Numba, LaTeX

From-scratch implementations of foundational quant methods to understand the mathematics. Options pricing (Monte Carlo, Greeks), portfolio optimization (Markowitz, risk parity), Fama-French factor decomposition. Textbook methods implemented for learning—not production tools.

Tech: Python, NumPy, SciPy, Pandas, Matplotlib

Learning project to explore C# and Unity game development. Built a third-person combat game to understand component-based architecture, real-time physics, animator state machines, and the Unity workflow. Implements dodge mechanics with i-frames, enemy AI with detection/pursuit states, and target lock-on system.

Tech: Unity 2022.3 LTS, C#, MonoBehaviour lifecycle, coroutines

Side-scrolling action platformer exploring real-time systems and software architecture. Implemented spatial partitioning for collision detection, event-driven game logic, FSMs for character states, enemy AI behaviors, and boss battles. Learning project focused on understanding OOP design patterns in practice.

Tech: Python, Pygame, OOP design patterns

Scientific Computing: Python (NumPy, SciPy, Numba), numerical methods (PDEs, ODEs, Monte Carlo), PyTorch
Machine Learning: Scikit-learn, feature engineering, TensorFlow/Keras, model evaluation
Backend & Web: FastAPI, PostgreSQL, Redis, SQLAlchemy, Celery
DevOps: Docker, Git, CI/CD pipelines
Quantum: Qiskit (HHL algorithm, circuit programming)
Desktop/GUI: PyQt6, real-time data visualization
Geospatial: PostGIS, GeoPandas, Plotly (3D visualization)
Other: LaTeX, SQL, TypeScript (basic)

I use AI coding tools (Copilot, Claude) to accelerate development—especially for boilerplate, documentation, and exploring unfamiliar APIs. The architecture, algorithms, and problem-solving are mine; AI handles the tedious parts.

• Email: adam.f.bentley@gmail.com
• Location: Wellington, New Zealand