The current polymer portfolio has been developed over the last 100 years, based on mostly empirical research on structures, properties, applications and processes. There is no time for a similar time-consuming exercise on CO2- or biobased polymers: 2050 is only 1 investment cycle away. Fortunately, machine learning (ML) principles have been applied to polymer science for the past decade. This so-called “Polymer Informatics (PI)” has large potential for the prediction of properties of new polymers and for the design of new polymers with desired properties. To unlock this potential, 3 challenges need to be addressed in parallel: 1) Data of polymers’ properties is scarce and fragmented. 2) Current ML-algorithms use oversimplified polymer fingerprints, lacking essential properties as 3D-morphology, molecular weight, co/mixed polymers, additives (e.g. stabilizers, pigments) and processing conditions. 3) Polymer design lacks integral assessment of viability, as expressed by chemical validity, safety (toxicity), techno-economic and sustainability. This needs a holistic assessment and the use of multicriteria decision making models to provide insights to researchers, product manufacturers and policy makers.

The computational study of complex polymeric materials is a very challenging field, due to the broad spectrum of the underlying length and time scales. We present a hierarchical multi-scale methodology for predicting the macroscopic properties of polymer-based nanostructured systems, which involves atomistic simulations, coarse-grained models, homogenization approaches and continuum models [1-3]. The CG and the continuum models are derived through a systematic “bottom-up” data-driven strategy, using information from the atomistic scale without any adjustable parameter [1]. Moreover, Deep Learning algorithms are developed to reintroduce atomic detail on the CG scale, and obtain atomistic configurations of long polymer chains [2].
The proposed computational methodology is applied to provide a fundamental understanding of the mechanism of mechanical reinforcement in glassy polymer nanocomposites, which is of paramount importance for their tailored design. We present a detailed investigation of the coupling between the density, structure and conformations of polymer chains with respect to their role in mechanical reinforcement [4-5]. We found that the effective mass density and rigidity of the matrix region change with filler volume fraction, while that of the interphase remains constant. The origin of mechanical reinforcement is attributed to the heterogeneous chain conformations in the vicinity of the nanoparticles, which involve a twofold mechanism. In the low-loading regime, the reinforcement comes mainly from a thin, single-molecule, 2D-like layer of adsorbed polymer segments on the nanoparticle, whereas in the high-loading regime, the reinforcement is dominated by the coupling between train and bridge conformations; the latter involves segments connecting neighboring nanoparticles

Polyurethane (PU) is a versatile polymer synthesized from a wide range of polyols and isocyanates, leading to a vast diversity in its chemical structure and properties. This variability presents significant challenges for efficient recycling, as different PU formulations exhibit distinct solubility characteristics, drastically influencing the post depolymerization separation, which is the largest contributor to the variance in economic feasibility. Addressing these challenges, this research integrates molecular dynamics (MD) simulations, machine learning (ML), and Aspen Plus (A+) process modeling to develop a systematic approach for PU recycling (Figure 1).

Key bottlenecks, including the lack of thermodynamic and physicochemical data for depolymerized products and the absence of optimized separation processes, are tackled through this framework. MD simulations predict critical properties such as solvation free energy, densities, heat capacities, partition coefficients, critical pressures, and temperatures for PU-derived molecules and mixtures.[1] Densities, heat capacities, and log(K)-values obtained from these simulations are validated through laboratory experiments and literature, showing errors below 5%. ML models, trained on both computational predictions and experimental data, extend the scope by rapidly estimating properties for unexplored compounds, significantly reducing computational costs.[2] Preliminary results using Gaussian process regression yield R² scores exceeding 0.946.

The validated data is then incorporated into Aspen Plus to model and optimize industrial-scale liquid-liquid separations with solvent recovery units, focusing on key parameters such as solvent selection, flow rates, temperatures and separation stages. By integrating MD, ML, and A+, this research enables a systematic evaluation and design of efficient, scalable, and sustainable PU recycling processes.

The global plastic waste crisis is one of the most pressing environmental challenges of our time. With millions of tons of plastic discarded annually, landfills, waterways, and marine ecosystems are increasingly burdened by non-degradable waste. As a result, the recycling and reuse of plastic have become imperative. However, current mechanical recycling methods often fail to fully separate different polymer types in recycled thermoplastics. These impurities can significantly affect the mechanical properties of the resulting materials, making it difficult to accurately predict their behavior using existing models and thereby limiting their reuse potential.
To address this challenge, this study proposes an innovative approach that integrates a neural network (NN) framework with a physically-based constitutive model to accurately predict the mechanical behavior of polymer blends with varying compositions. Rather than directly fitting the stress-strain response, our approach focuses on predicting the athermal shear resistance, an essential internal scalar variable derived from the physical model. By doing so, the NN-based constitutive model reduces reliance on extensive experimental data, simplifies the parameter identification process and improves its overall generalization capabilities. We validate this model by applying it to various polymer blends, where the predicted results closely match experimental data. This demonstrates the effectiveness of the NN-based constitutive model in capturing the mechanical response of different polymer compositions and highlights its potential for optimizing recycled polymer blends across a wide range of applications.

Kinetic simulations of radio-oxidation in polymers follows schemes developed several decades ago; recent updates on the list of relevant mechanisms as well as corresponding rates are mostly indirectly extracted from experiments [1]. In this paper we describe some recent advances in the investigation of atomic scale mechanisms relevant for aliphatic polymer oxidation taking the example of polyethylene. Based on a polymer model manageable by first principles calculation but still containing the main features of a semi-crystalline polymer, we determine relevant energy barriers and we corroborate our findings with first principles molecular dynamics simulations. Our calculations are based on density functional theory with a van der Waals exchange-correlation functional and, in some cases, we resort to a hybrid functional for comparison.
After verifying the main reactions involved in the standard basic oxidation scheme [1], in particular the formation and decomposition of hydroperoxides [2], we investigate several reactions involving alkoxy radicals, which can originate from bimolecular reactions between peroxy radicals [3]. The results clearly show the crucial role of this radical in the whole radio-oxidation kinetic path.
Furthermore, we consider the radical scavenging capacities of phenolic antioxidants, taking butylated hydroxy-toluene (BHT) as a prototype. Using static and dynamic simulations, we highlight the fact that BHT can easily eliminate alkoxy radicals and could possibly influence the concentration of peroxy radicals, but is not able to scavenge the primary alkyl radicals which trigger the initiation of radio-oxidation [4].

Polymers are “products by process” in which the final polymer properties, such as the molecular weight distribution and macromolecular architecture, are determined during the reaction. Polymerization of acrylic monomers is complicated by numerous side reactions, making understanding and controlling these processes highly challenging. In this work, the use of machine learning to develop data-driven predictive models of acrylic monomer emulsion polymerization, and how these models enable control over the final polymer properties, is discussed.

In order to develop accurate data-driven predictive models using only limited datasets, we demonstrate the use of Polymer Chemistry Informed Neural Networks (PCINNs), which incorporate fundamental polymerization kinetics into neural network training. The PCINN approach is used to develop predictive models of the emulsion polymerization of poly(butyl acrylate). This hybrid approach enables accurate predictions of molecular weight distributions with minimal experimental data, while maintaining the speed advantages of neural networks.

Building on these predictive models, we applied reinforcement learning (RL) to develop control strategies that optimize process conditions. These RL approaches successfully achieved target molecular weight distributions even in the presence of process disturbances, demonstrating robust performance in real-world conditions.

These complementary approaches represent a promising direction for polymer reaction engineering by combining domain knowledge with machine learning to overcome traditional computational and experimental limitations.