The cosmetic, food and nutrition, pharmaceutical, chemical and materials industries face recurring tensions during the product design phases.
The challenge: innovate / design high-performance, competitive, stable, eco-sustainable products, on short notice: Time to Product.
In addition, there are aspects such as process traceability, collaboration management, knowledge historization, control of R&D costs ... As examples, we can cite the integration of a new innovative active in a formula, the development of a food recipe by minimizing the rate preservatives, the replacement of a raw material subject to regulations, the implementation of an industrial process with less energy footprint ...
Technicians, formulators, engineers, scientists strive to push the boundaries of innovation by experimenting with processes, settings and mixtures of materials. The intuitive approach consists of starting with a "first guess", i.e. a process / formula approaching the target, then by a set of increments progressive inputs, to navigate visually to try to achieve the expected performance. This process is tedious, because for a formula comprising 10 parameters each taking 10 values, there are 10 billion combinations. Which is prohibitive for physical experiments.
In general, the expertise, the operational constraints, the business know-how make it possible to avoid certain improbable combinations for reduce the space of possibilities to a few hundred or even thousands of configurations. But it remains high all the same. In addition, since the exploration is not planned, it becomes difficult to learn the lessons of the causal relationships. within the process, to capitalize on the tests carried out. This can lead to redoing the same experiences for a slightly different need.
Mathematical models can be used through digital simulation and therefore virtualization of experiments. However, this assumes that the underlying model linking inputs to outputs is known, requiring years of research. to establish and implement it. This model lapses as soon as a new component is added or removed. Which is not possible for an industry that wants to be agile.
Statisticians invented the methodology of experimental designs (Fisher 1935). In summary, it consists of planning a limited set of experiments intended to crisscross the field of research, then to establish with these experiments a simplified model of the relation of cause and effect (linear, quadratic, first order interactions…). The inference on the objective is made by exploiting the coefficients of the regressions, and predictions. This methodology has shown its effectiveness in various fields in the past. However, there are 2 drawbacks to remember:
- you must have a statistical background to implement them.
- the current complexity of formulas and procedures can no longer be explained simply with the over-simplified assumptions of classical experimental designs.
In chemistry, for example, we will have phase transitions, breaks in phenomenon ... A daily example is evolution the temperature of a block of ice subjected to a heat source. Temperature levels are observed during changes of state, and almost linear variations, for the homogeneous phases, giving way to a stepped model.
It is to these challenges that InFLOWS offers an answer with alternatives based on intelligence
artificial in particular Machine Learning). Machine Learning is a branch of artificial intelligence that is based on approaches
statistics to give computers the ability to "learn" from examples. In particular, to model relations without having to explicitly
specify or program them. Very well, this is what we are trying to model the behavior of substances without an explicit postulate
a priori on their synergies and then generalize the hidden relationship between the inputs and performance of a product.
An example of a model: make a program that knows how to differentiate an apple from a pear.
We can see that this is an easy exercise for a child, but more difficult to specify explicit rules for a computer program.
For this example, Machine Learning will consist in presenting to a computer program inspired by the structure of the brain, successively several examples of apple and telling him this is an apple and several examples of pear. At the end of this automated learning, the program will be able to extract by itself the hidden characteristics (eg: invariants) that make an apple an apple and vice versa. The program will then be able for a new image to say which fruit it corresponds to, this is the faculty of generalization. You will understand, this example is greatly simplified, the process of architecture and learning of Machine Learning methods is much more complex and is one of the research topics of greatest interest. Their implementation requires data science experts (Data Scientist), to implement the analysis pipeline either on fresh data or for retro-analysis on historical data. These profiles are not always available, and can represent a point of contention.
To meet these challenges, InFLOWS has surrounded itself with a multidisciplinary team to cover the entire development cycle digital assistants to give back autonomy to scientists (physicist, chemist, computer scientist, developer, administrators, ergonomists, UI-UX, data engineer, data scientist, Agile project manager etc.)
InFLOWS provides a catalog of turnkey applications to boost process engineering. The ergonomics, interfaces, onboard calculation engines are the result of more than 10 years of experience by machine learning experts. The procedures are automated so as to reproduce the steps of the data scientist workflow, and to provide feedback interfaces and visualization for researchers. Our assistants are deployed in the internal IS of our customers with a historization on-premise database experiments. The projects carried out are digitized and their results can be shared within the company offering the possibility of collaborating with distant colleagues. Monitoring interfaces make it possible to monitor project KPIs, and application administration. We also offer the development of customized applications for the needs of our customers. Our working methods are largely inspired by Agile methodologies such as Design Sprint, or how to prototype in 5 days.
Smart Labs : Where AI meets Process Engineering !