Two modeling approaches dominate the scene in product manufacturing. The so-called mechanistic models (also referred to as ab-initio or deterministic) had been the only choice for decades, are
driven by process knowledge and focus on a precise mathematical description of the process chemico-physical phenomena. In the last decade, the focus shifted towards the so-called machine-learning
models, which are driven by data and consist in sophisticated algorithms to extract patterns and correlations among large and complex sets of data.
To realize our vision, we demand our models to be flexible and inexpensive. Flexibility means having models that can be easily applied to different products and to different
equipment, scale or process conditions without any loss of knowledge and with zero or minimal experimental effort. Inexpensiveness means having models that are easy to set up, that requires no or
minimal user intervention to manage data, and that requires very little and standard data and experimental effort to be trained, while providing the desired accuracy in prediction. Our hybrid
technology is the sole technology allowing to integrate both these two fundamental characteristics.
Machine Learning MODELS
Comparing modelling technologies
Hybrid models fuse mechanistic and machine learning models and integrate process knowledge and data information, causality and correlation.
Hybrid models are simple and fast to establish and do not require dedicated experiments
Hybrid models retain most of the predictability of mechanistic models
Hybrid models retain the model flexibility of deterministic models and the ability to adapt to different process configurations
Hybrid models can handle large data and variables for which non-specific knowledge is available
Hybrid model integrate the ability of ML model to find complex correlations and patterns among inputs and can be used to transfer such data-driven knowledge
Mechanistic models are based on process knowledge, focus on the chemico-physical description of the process and on the definition of causalities relations among variables.
Once the model is established, this can be applied to any process configuration and knowledge can be transferred to new products with few dedicated experiments
Extrapolation to new process conditions and ranges of the parameters is typically reliable and robust
The estimation of the physical parameters of the model often requires dedicated non-standard experiments
Knowledge to build the model is very often not available for complex variables and a long work to select and proof models and suitable approximations is needed
Machine Learning Models
Machine learning (ML) models are based on data information, focus on the definition of multi-variate correlations and patterns among variables.
ML models are simple and fast to establish and do not require dedicated experiments
ML models are extremely powerful in finding correlation among parameters for which no specific knowledge is available, making them ideal to identify and transfer non-specific
ML models are rigid in the input variables and in the data structure
They are difficult to adapt to different process configurations and set-ups
They lack robustness in model extrapolation and scenario analysisd scenario analysis