Gurpreet Singh

Experience

• 

  Center For Subsurface Modeling

  Dec 2014 - Dec 2018

  1. Developed a space-time domain decomposition approach with a time-concurrent, parallel solution algorithm for multiphase flow and reactive transport problems in porous composites. The prototype implementation was verified and benchmarked as a stand-alone Matlab framework with direct compressed row storage to promote faster developments on limited main memory hardware. The time-concurrent scheme eliminates time-dimension induced sequential instructions bottle-necking the parallel performance of a compute framework. This approach is now being expanded for general use in problems with strong ordering or sequence requirements.2. Developed and implemented (Matlab and Python prototype) model order reduction (dimension-reduction) techniques to reduce compute cost for practical problems of interest in subsurface porous media with minimal errors (user-specified tolerance) in quantities of interests either chosen from the implemented library or user specified.3. Developed and implemented an approximate Jacobian non-linear solver that outperforms linear solvers with conventional physics informed preconditioners. The solver was implemented and benchmarked in IPARS with reported wall-clock speedups of 1.25 to 4x compared to commercial and non-commercial off the shelf solver libraries. Show less 1. Developed and implemented an adaptive homogenization approach for upscaling heterogeneous porous medium for computational efficiency (4 – 20x wall-clock speedup) and accuracy. Previously implemented and benchmarked as a Matlab prototype, this approach is currently being ported to Tensorflow-GPU paradigm as a physics informed (mass and concentration conserving) dimension reduction technique for faster computations and to promote ease of use.2. Developed a computationally efficient and accurate framework for subsurface reactive flow and transport processes (stand-alone Matlab prototype). The framework uses the fact that an equilibrium reaction is a kinetic reaction in a finite time limit to bring the otherwise incongruent concepts to the same frame of reference. 3. Extended the parallel framework in IPARS (Integrated Parallel Accurate Reservoir Simulator) to a higher-order mixed finite element scheme (developed earlier as an internal module/library) for general use in simulating and predicting an array of subsurface flow and transport processes. Show less

  • Research Associate

    May 2017 - Dec 2018

  • Postdoctoral Research Fellow

    Dec 2014 - May 2017
• 

  Computational Hydraulics Group

  Dec 2018 - Aug 2020

  Research Scientist

  1. SCA-Net: Developed a self-correcting autoencoder using bi-orthogonal representations for hyper-spectral feature extraction that outperforms SOTA results on benchmark problems. SCA-Net has the lowest main memory requirement and highest accuracy compared to all the available commercial and non-commercial implementations. The framework is implemented with Tensorflow backend and Keras frontend to use efficient GPU computations. This formulation has strong ties to Shannon’s information entropy and is being expanded for automated data compression under user restrictions with separable encoder and decoder. SCA is fully interpretable where the learned weights can be verified by the user for network diagnosis.2. Range-Net: Developed a streaming singular value decomposition that outperforms SOTA randomized algorithms. Range-Net has a low memory requirement for compute efficient and accurate ranking of important samples and features in Big Data. The prototype was implemented with Tensorflow backend and Keras frontend and is currently being modified for fast and interpretable run-time anomaly identification and basis set augmentation. Range-Net can be used as a non-invasive, neural network diagnosis tool without interrupting (parallel instructions) an existing instruction graph.3. HNPF: Developed a Hybrid Neural Pareto Filter approach for high-dimensional problems with multiple competing objectives and multiple viable solutions (different solutions with the same energy metric at scale). HNPF gathers and parametrizes the solutions for easy visual assessment and decision making. The double gradient descent implementation (Scalable Unidirectional HNPF) uses Tensorflow-GPU libraries for an inner iteration with the outer iteration made available as an interface for specifying objectives rendering direct control to the user. Show less

• 

  The University of Texas at Austin

  Aug 2019 - Apr 2021

  Adjunct Faculty Instructor

  -Computational Engineering, Aerospace & Engineering Mechanics-Introduction to Numerical Methods, Biomedical Engineering

• 

  Resermine Inc.

  May 2021 - now

  1. Transition trusted algorithms and workflows to Tensorflow utilizing GPUs for fast and accurate compute solutions. Assist energy industry digital transformation from on-premises computations to cloud compute infrastructures.2. Converted CRM a, widely used model for oil field predictions, to Neural-CRM reducing runtime from hours to less than a minute.

  • Sr. AI/ML Development Consultant

    Feb 2023 - now

  • Algorithms Consultant

    May 2021 - Feb 2023