WHITELEY'S

Impeller Mechanics with Physics-Informed NN's

GIF 1 GIF 2
  • Physics-Informed Neural Networks (PINNs) learns the rotational frame Navier Stokes equation in three dimensions
  • A PyTorch based transformer neural network trained on Lambda Labs NVIDIA H100 Tensor Core clusters
  • CFD data generated by sampling design variable parameters via latin hypercube sampling
  • Sample data fed to automated meshing script which feeds to automated custom OpenFOAM CFD C++ solver for full "real" flow fields
  • Data is used for unsupervised continuous learning of entire non-linear flow fields in the 3D domain for any impeller design
  • Usecases:
    • A fast surrogate to replace CFD in the loop of an evolutionary optimisation algorithm, allowing fast, dynamic and adaptable design procedures
    • A model that can be used as a "digital twin", where live data is fed to a model and predicts physical behaviour before unforeseen circumstances

Technical Analysis with Deep LSTM RNN's

GIF 1
  • Developed a Multi-layered, Multi-dimensional LSTM Recurrent Neural Network model to predict stock prices and volume
  • Integrated multiple technical indicators such as Moving Averages, RSI, and MACD as input features for enhanced predictive accuracy, regularised data using appropriate transformations
  • A PyTorch based unrolled RNN neural network trained on Lambda Labs NVIDIA H100 Tensor Core clusters

Rocket Guidance with Non-Linear Control

GIF 1
  • Developed an airbrake mechanism that extends flaps to slow a rocket down to desired apogee.
  • Utilized KiCad for comprehensive circuit design of a flight computer, including schematics and PCB layout. Combining different components and sensors (IMU, Baro, GPS, Radio)
  • Programmed the flight computer and GNC in C and C++, designed a state of the art Extended Statistical Kalman Filter for precise state estimation, using quaternion mathematics for orientation and position calculations.
  • Implemented PID control by building lookup tables of drag coefficients using OpenFOAM CFD automations in C++, which are called by the rocket control system during flight.
  • Calculated the predicted apogee in real-time using an efficient Runge-Kutta numerical algorithm, enabling the solution of a non-linear ODE that accounts for rocket drag.

Tomas Whiteley

MEng Mechanical Engineering

University of California, Santa Barbara
University College London
github | linkedin