Science / the algorithm

How STOCHOS works.

STOCHOS is built on DIM-GP, a Deep Infinite Mixture of Gaussian Processes. The deep network does not make the prediction. It learns how your data behaves at each point and hands that to a Gaussian process, which answers with a confidence range.

Request a Demo Explore STOCHOS

DIM-GP combines the capacity of deep learning with the uncertainty estimates of Gaussian processes in one non-stationary probabilistic model.

The gap

One smoothness does not fit a real design space

Real responses change character: smooth in one region, sharp in another, densely tested here, barely sampled there. A standard model picks one smoothness for all of it. DIM-GP adapts locally.

Classical GP

Assumes one smoothness everywhere, and slows sharply as data grows.

Deep networks

Scale to complex relationships, but return one number with no confidence.

DIM-GP

Learns smoothness and noise locally, and puts a confidence range on every prediction.

Noisy data

It does not chase outliers

Measurements scatter, batches vary, solvers leave numerical noise. DIM-GP learns the noise level point by point, so it separates trend from scatter and does not bend the fit toward an outlier.

A standard model fits a single curve through noisy data with no confidence bounds; it warps toward an outlier and gives no indication of how much to trust the prediction. — Standard model. One curve, no confidence. It trusts the noisy point blindly.

DIM-GP returns a mean prediction with one, two, and three sigma confidence bands that stay tight where data exists and widen where the model extrapolates, and discounts the noisy outlier. — DIM-GP. Mean plus confidence bands. Confidence widens where data is missing.

The full picture

What comes with the model

01Uncertainty

Built in, not bolted on

Every prediction carries a confidence interval, local to that point. A single accuracy score hides the weak spots.

02Small data

Built for data that is expensive

When a data point costs hours or money, big-data methods starve. DIM-GP takes the trend a small set of runs supports, and the interval says where the next run should go.

03Data shapes

One approach, every data shape

One probabilistic core for scalars, signals, images, 3D fields, meshes, and point clouds, plus multi-fidelity variants. 14 DIM-GP models ship in STOCHOS Flow.

04Optimization

From model to decision

STOCHOS feeds the interval to a Bayesian optimizer that adapts after each result and returns a Pareto front when objectives compete. How Bayesian optimization works

05Ongoing research

The algorithm is not standing still

New model families, faster training, new data types in regular releases. The team that develops the algorithm is the team you talk to.

Methods Surrogate modeling Uncertainty quantification Multi-fidelity modeling Sensitivity analysis Generative design

See DIM-GP on your own data

Request a demo and we will show STOCHOS and STOCHOS Flow on a problem close to yours.

Request a Demo Back to Science

Common questions

01What is DIM-GP?

DIM-GP, a Deep Infinite Mixture of Gaussian Processes, is the core STOCHOS algorithm. It combines neural networks and Gaussian processes into a non-stationary, probabilistic model designed for small, noisy datasets, changing geometries and meshes, and multi-fidelity data, without manual hyperparameter tuning.

02Why combine deep learning and Gaussian processes?

Deep learning brings capacity to model complex, high-dimensional relationships. Gaussian processes bring uncertainty estimates. Engineering needs both: a flexible model that also reports how far to trust each prediction. DIM-GP delivers both in one model, and trains several covariance functions at once, letting the data decide which ones carry the model, so the tuning step that normally needs an expert disappears.

03How is DIM-GP different from a standard Gaussian process?

A standard Gaussian process assumes the same smoothness everywhere and becomes slow as datasets grow. In DIM-GP, a deep network sets the process parameters locally, so smoothness and noise adapt across the design space, and training runs in batches, which keeps large datasets practical.

04Can STOCHOS run on-premise?

Yes. STOCHOS runs on Windows and Linux, on CPU or GPU, as a Python 3.12 package built on PyTorch, and can be installed without an internet connection by providing the packages directly. Core modeling and optimization run on your own infrastructure, so sensitive simulation, test, and formulation data stays under your control.