What is Under the Hood

gpCAM is an advanced Gaussian process tool, combined with HPC function optimization. 

gpCAM's power stems from its flexibility: By design, everything that is mathematically possible with GPs is possible in gpCAM.  The user can limitlessly customize most parts of the Gaussian process as they become more familiar with the underlying mathematics. One could imagine a car engine; the motor block is the core code of gpCAM, all other parts that make a well-working car are provided, but they can be exchanged to create the perfect vehicle for a particular purpose. Some of the building blocks that can be defined by the user and imported into gpCAM are:

• Any kernel function to constrain the set of model functions, for extra flexibility, and more accurate uncertainty quantification.

• Any parametric prior-mean function to encapsulate a physics-based model.

• An optimizer that replaces the standard optimizers, for instance, for constrained training.

• A flexible block-MCMC for advanced fully Bayesian training.

• Any acquisition function to inject patterns and characteristics of interest.

• A cost function is used to make sure measured points will minimize costs while maximizing knowledge gain.

• In addition, gpCAM is entirely agnostic to the kind of input space: gene sequences, molecular descriptors, integers, or the traditional Euclidean space. 

All this flexibility means that gpCAM naturally includes the ability for physics awareness and also multi-task learning.

Torch and DASK-based high-performance computing mean that gpCAM can take full advantage of supercomputers.

Read More Here:

• Compactly-supported nonstationary kernels for computing exact Gaussian processes on big data

• Exact Gaussian processes for massive datasets via non-stationary sparsity-discovering kernels

• A unifying perspective on non-stationary kernels for deeper Gaussian processes

• Advanced Stationary and Non-Stationary Kernel Designs for Domain-Aware Gaussian Processes

• High-Performance Hybrid-Global-Deflated-Local Optimization with Applications to Active Learning

• Gaussian processes for autonomous data acquisition at large-scale synchrotron and neutron facilities

• Autonomous materials discovery driven by Gaussian process regression with inhomogeneous measurement noise and anisotropic kernels

• Hybrid genetic deflated Newton method for global optimisation

• Advances in Kriging-Based Autonomous X-Ray Scattering Experiments

• A Kriging-Based Approach to Autonomous Experimentation with Applications to X-Ray Scattering

Many of the recent features have not been published yet. The papers will be linked here soon.