PI Probaligence GmbH, an innovative company in probabilistic AI solutions and partner in the CADFEM Group, has developed an advanced set of machi-ne learning tools called STOCHOS©, which uses a probabilistic machine learning algorithm approach. STOCHOS© performs sample- and resource-efficient, expert-independent probabilistic machine learning, using both existing data and planning new data collection. STOCHOS© can learn from all kinds of data shapes and formats, e.g. transient 2D or 3D FEM / CFD and re-place expensive simulations like, for example, aerodynamic simulations. Even signal data like in NVH tasks are possible. It does not matter whether the data comes from real experiments or simulations, showing its high flexibility.
STOCHOS© supports all three kinds of machine learning, supervised, unsupervised and reinforcement learning. In supervised learning, it is excellent for replacing simulation solvers and performing optimization. Un-supervised learning would be for example predictive maintenance tasks where the model analyzes anomalies in sensor data. Reinforcement learning tasks cover all kinds of real time control tasks like in robotics or for example autonomous driving. The software works on various hardware platforms, including CPU, GPU, and microcontrollers, allowing for real-time responses.
A key feature of STOCHOS© is its multi-fidelity computation capability, which increases computational efficiency or handles very complex problems with limited computational resources. It allows to combine different fidelity levels of data sources like for example coarse and fi ne mesh simulation results or to combine experimental and simulation data within one model to solve simulation calibration tasks. This makes STOCHOS© exceptionally good at managing many different simulations needs without losing performance or accuracy.
PI Probaligence’s STOCHOS© software combines the best features of Deep Learning (DL) and Gaussian Processes (GP) while avoiding their downsides. Unlike large ML companies that often rely on deep learning and need extensive hyperparameter tuning to fi nd the best network setup, STOCHOS© does not need any kind of hyperparameter training. In addition, the data stays securely within the customer’s premises, avoiding the need for expensive and resource-heavy cloud computing solutions.