Data-based engineering science and technology of polymers and filled polymers – Data-BEST

The initial objectives of the project were multiple:
(i) Develop machine learning techniques capable of operating within the limits of sparse data;
(ii) Data-driven modeling of the conformational evolution of molecular chains and fibers in concentrated suspensions calculated by physics-based simulations at very fine scales;
(iii) Description of complex microstructures (foams) with concise and accurate descriptors operating directly on microstructure image data;
(iv) Process descriptions by linking properties and performance with input parameters concerning materials and processes: application to reactive extrusion;
(v) ??Hybrid modeling by expressing the data as a first-order model (knowledge or physical model) and a deviation expressed from the data by machine learning.

Returning to the objectives outlined above, below we describe the main research actions that resulted from this:
(i) Development of machine learning techniques capable of operating in the low-data limit. Two techniques were proposed: Code2Vect and iDMD. These two codes enabled classification and regression to operate in the low-data limit.
(ii) Data-driven modeling of the conformational evolution of molecular chains and fibers in concentrated suspensions, calculated by physics-based simulations at very fine scales. The use of neural networks for molecules and PCA/Code2Vector for suspensions allows for scaling, where fine simulations produce the data for learning models at subsequent scales.
(iii) Description of complex microstructures (foams) with concise and precise descriptors operating directly on microstructure image data. TDA (Topological Data Analysis) was used to describe microstructures and create regressions based on homogenized properties.
(iv) Process description by linking properties and performance with material and process parameters: application to the reactive extrusion process. Various machine learning techniques were successfully used.
(v) Hybrid modeling by expressing the data as a first-order model (knowledge or physical model) and a deviation expressed from the data using machine learning.

Polymers and viscoelastic fluids can be described at various scales, but precise molecular-level descriptions are expensive and impractical for industrial process simulations. Artificial intelligence (AI) offers a solution for efficiently scaling back and forth. Machine learning techniques, such as neural networks, have been used to predict the rate of molecular conformation change from microscopic simulations, based on the conformation and local flow properties.

AI has also brought advances to the polymer industry, particularly in the reinforcement of polypropylene with a thermoset phase. A project with TotalEnergies, based on a Hutchinson patent, studied the dispersion of a thermoset phase in a polypropylene matrix using in situ polymerization. Several polypropylenes and epoxy resins were tested to improve the mechanical properties of the thermoplastic polymer. A polypropylene grafted with maleic anhydride was used as a compatibilizer.

Conventional simulations are not ideally suited to describe polymer blends, especially those with low viscosity ratios. Data science and machine learning can improve these simulations. A hybrid approach, combining data and models, was tested. The data came from simulations (Ludovic software) and experimental measurements, with four outputs to be predicted: torque, pressure, motor power, and outlet temperature. The simulation data has two inputs (flow rate and rotational speed), while the experimental data includes an additional parameter (dispersed phase).

The experimental dataset contains 47 samples, divided into 38 points for regression and 9 points for verification. The procedure consists of subtracting the first-order model values ??from those of the ground truth and applying a regression to this deviation. This approach facilitates the construction of the regression and improves its performance, making the deviation model easier to determine than the model based exclusively on the data.

Simulation under process conditions, as well as its validation, for polymers and reinforced polymers was abandoned due to the lack of resources to collect data, to focus on reactive extrusion. The abandoned tasks are being progressed within the framework of the ANR Bancodemm which took over from DataBEST.

Ibañez, R.; Castéran, F.; Argerich, C.; Ghnatios, C.; Hascoet, N.; Ammar, A.; Cassagnau, P.; Chinesta, F. Data-Driven Modeling of Reactive Extrusion. Fluids 2020, 5, 94.

Castéran, F.; Delage, K.; Cassagnau, Ph.; Ibanez, R.; Argerich, C.; Chinesta, F. Application of Machine Learning tools for the improvement of reactive extrusion simulation. Macromol. Mater. Eng. 2020, 305, 2000375.

Castéran, F.; Delage, K.; Hascoët, N.; Ammar, A.; Chinesta, F.; Cassagnau, P. Data-Driven Modelling of Polyethylene recycling under High Temperature Extrusion. Polymers 2022, 14, 800.

Castéran, F.; Delage, K.; Cassagnau, Ph.; Chinesta, F.; Cueto, E.; Ammar, A.; Garois, N. L’Intelligence Artificielle appliquée à la mise en œuvre réactive des polymères. L'actualité chimique. N°456, 457, 458, Novembre-Décembre-Janvier 2020-2021.

The polymer industry is typically facing these new challenges for the production of new materials with new functional and integrated properties in different domains of applications. Nowadays, although new technologies are emerging (3D printing) polymer processing is still based on conventional approaches in terms of modeling, simulation and scale-up for the production. So, can we imagine new original and realistic ways based on the recent developments on artificial intelligence ?

Traditionally, when a material in its in-service conditions was considered, a constitutive equation was selected (among different possibilities) guided by the accumulated experience and/or the experimental observations, and then calibrated in order to identify all the parameters that it contains. This procedure faces two main difficulties: (i) In general establishing a constitutive equation needs time and the intuition and accumulated knowledge of experienced specialists; and (ii) Even when accurate constitutive equations could be derived by spending a sufficient material and intellectual effort, the recent explosion of material proposals requires a new procedure to describe material properties in its in-service conditions easier and faster.

In the opinion of the applicants, Industry 4.0 needs a change of paradigm in the description of materials and processes as just described. It is also important to note that here we are not replacing or claiming the replacement of experienced specialists, because deciding the data to be collected, the one that describes the state of the material in the process, etc … needs a strong and mature experience and knowledge, that is DATA-DRIVEN IS NOT ONLY DATA, IT IS DATA IN THE RIGHT HANDS, and only in these circumstances data become information to finish as knowledge.

The biggest challenge and main originality that the present research project proposal embraces, could then be formulated as follows:

- Can simulation proceed directly from data by circumventing the necessity of establishing a constitutive model in the traditional sense? Of course balance equations are kept, whereas the effort in establishing mathematical expressions of (too) complex constitutive equations is relaxed.
- How data-based modeling can enrich the establishment of the models (viewed as mathematical objects) describing the behavior? It is important to note that the main advantage of models is their capability of predicting outside the domain that served for their establishment (the so-called extrapolation) where data-based modeling encounter difficulties to make the job.
- How models can help data to be consistent with principles (convexity of potentials, …), filtering measurements noise, determine the quantities to be measured (electrical conductivity, shear and elongational viscosities, tensile-strength properties, ….) or announcing the necessity of using internal variables, …
The main project tasks/objectives to be accomplished are:

- To propose data-based model learners able to discriminate conservative and dissipative behaviors in the case of polymers and reinforced polymers;
- To couple model learners with rheometers for simple flows;
- To define strategies for collecting data from more complex flows exhibiting shear and elongational behaviors;
- To describe the mechanisms of mixing and dispersions of a minor phase in terms of reinforcement and specific properties (electromagnetic shielding for example)
- To address the issue related to objectivity (frame indifference) in a data-driven framework;
- To verify the data-driven approach by emulating the real scenario from model-based simulations;
- To validate the approach at the laboratory scale and at the different scales of processing.

These objectives will be achieved in the framework of the developments of new materials that constitutes a major advantage of data-driven approaches:
- Polymer melt processing
- Reactive polymers extrusion
- Filled polymers with conductive fillers