Conference Agenda

Identifying accurate kinetic models for new biochemical systems is a great challenge. Kinetic models are represented by differential equations, where the parameters and terms of the symbolic expressions hold physical significance. Hybrid modelling and traditional machine transfer learning can leverage previously discovered relations about different but related systems to minimise the time and experimental resources necessary to develop accurate predictive models for new systems. However, these methods are non-interpretable, by only updating the data-driven and kinetic parameters of an existing hybrid model, they leave the kinetic model structure unchanged so are unable to provide additional physical insight into the newly investigated system.

To address this challenge, we propose a novel model structural transfer learning methodology that integrates symbolic regression (SR) with artificial neural network (ANN) feature attribution to streamline the discovery of interpretable differential equations for new biochemical reaction systems. The feature attribution technique effectively guides SR towards targeted modifications for existing erroneous or low-fidelity mechanistic models, addressing the traditional challenge of efficiently exploring the large combinatorial space of expression structures. More importantly, this approach, combined with strategic sampling and model-based design of experiments (MbDoE), maximises knowledge extraction while minimising experimental resource requirements.

Through a comprehensive in-silico case study, our framework effectively adapted the structure of a kinetic model taken from one biochemical system for a new but related biochemical system, discovering the underlying kinetic equations. The predictive accuracy and uncertainty are then benchmarked against traditional hybrid modelling techniques. The impact of prior knowledge quantity and fidelity are also explored, demonstrating the framework’s ability to either rebuild equations from scratch or make targeted corrections during model structural transfer learning. To glean a physical interpretation of the differences in the underlying process mechanisms, the terms that have been modified could be compared and their physical meaning could be easily understood, altogether highlighting the framework’s significant potential for advancing automated knowledge discovery and novel biochemical process development.

The CompArt project explores new approaches to modelling complex systems, focusing on the integration of next-generation compartmental models powered by artificial intelligence (AI). Our primary objective is to streamline three-dimensional (3D) Computational Fluid Dynamics (CFD) simulations while preserving critical spatial characteristics and system heterogeneity. By modelling the 3D system as a network of interconnected sub-systems with uniform properties^[1], we significantly reduce computational costs and enhance simulation efficiency without compromising accuracy. A central innovation of CompArt lies in the application of AI-driven clustering techniques for the compartmentalization process. This approach enables the AI to autonomously determine the optimal number of compartments based on user-defined parameters, automating a traditionally expert-driven process.
Consequently, our framework becomes accessible to a broader audience, including non-experts, while delivering transparent and user-friendly outputs. Although multiple clustering techniques such as k-means, agglomerative clustering, and DBSCAN were tested, here we report the results obtained using Self-Organizing Maps (SOM). SOM are unsupervised neural networks that project high-dimensional data onto a two-dimensional grid while preserving the topological relationships of the input space. This makes them particularly effective for clustering spatial data, such as velocity distributions. The results demonstrate how the choice of input parameters influences clustering. When only the velocity distribution is used, SOM accurately capture zones of differing velocities, mapping system heterogeneity effectively. However, relying solely on velocity is insufficient in many chemical engineering processes. In reactive crystallization, for instance, a critical variable is the turbulent kinetic energy dissipation rate (ε)^[2]. When both velocity and ε are used as inputs, the optimal number of clusters decreases, reflecting a trade-off between the two variables that better describe the process heterogeneity. The key takeaway is that the algorithm automatically determines the optimal hyperparameters, such as the map size and learning rate, through a Silhouette score, removing the need for user intervention or prior knowledge. Additionally, the algorithm can handle multiple controlling variables simultaneously. Beyond velocity and ε, it could incorporate variables such as the saturation field or temperature, critical parameters in processes like pharmaceutical manufacturing. In this regard, CompArt aims to address the critical challenges of scaling processes, particularly where process heterogeneity impacts Critical Quality Attributes (CQA). This capability is vital in pharmaceutical manufacturing applications such as stirred tank reactors, where scaling decisions directly influence product quality. The AI-driven model, capable of integrating key variables, represents a flexible and powerful tool for capturing system heterogeneity, enabling more efficient and accurate process simulations across various industrial sectors.

References

(1) Jourdan, N.; Neveux, T.; Potier, O.; Kanniche, M.; Wicks, J.; Nopens, I.; Rehman, U.; Le Moullec, Y.; "Compartmental Modelling in chemical engineering: A critical review", Chemical Engineering Science, 2019, 210, 115196.

(2) Raponi, A.; Achermann, R.; Romano, S.; Trespi, S.; Mazzotti, M.; Cipollina, A.; Buffo, A.; Vanni, M.; Marchisio, D.; "Population balance modelling of magnesium hydroxide precipitation: Full validation on different reactor configurations", Chemical Engineering Journal, 2023, 477, 146540.

With climate change taking a prominent role in driving innovation, focus on developing novel renewable energy technologies has become imperative. Recently, perovskite solar cells have achieved an efficiency of>26.5%¹ in single-junction devices and >34.5%¹ in silicon/perovskite tandem solar cells. However, issues related to stability² still pose the most pressing challenge in taking this novel technology to market which has the potential to compete with existing PV technologies in terms of cost and environmental footprints³. With the huge amount of data in the literature on perovskite solar cells, it is crucial to use this data optimally and efficiently to not only identify the trends in the existing data but also to generate new recipes that might have better stability than the existing recipes in the literature. Moreover, the advent of large language models (LLMs) and significant progress in using them for specific tasks have opened the venues to test their abilities in dealing with more scientific problems apart from the more generic tasks like summarization or text completion⁴. In this research, an attempt has been made to use both open-source LLMs like Llama and closed-source LLMs like OpenAI GPT models to generate recipes that could have similar or better performance compared to the existing ones. This involves testing different kinds of data formats, large language models, and hyperparameter fine-tuning, to get the best performance using these models. The modeling approach is a first attempt to generate not only material choices for main layers like the electron transport layer, hole transport layer, perovskite layer, and rear and front electrodes but also the different additives, processing routes, and other process steps involved in the complete fabrication of the perovskite solar cells. The detailed architecture in dealing with this problem involves using LLMs for both the generation of the strings and then the prediction of the generated strings to fine-tune the generating LLM to achieve the required target of stabilities. The architecture is inspired by the framework of tuning the instruction-tuned GPT models. Using the proposed framework, the capabilities of LLM models in interpreting the causal relationships will help in tackling the challenges of optimizing material-process design problems (MPDPs) along with identifying the limitations of the LLMs.

References:

1. National Renewable Energy Laboratory, Best Research-Cell Efficiency Chart
2. Duan, L., Walter, D., Chang, N., Bullock, J., Kang, D., Phang, S. P., ... & Shen, H. (2023). Stability challenges for the commercialization of perovskite–silicon tandem solar cells. Nature Reviews Materials, 8(4), 261-281.
3. Bhati, N., Nazeeruddin, M. K., & Maréchal, F. (2024). Environmental impacts as the key objectives for perovskite solar cells optimization. Energy, 299, 131492.
4. Xie, T., Wan, Y., Zhou, Y., Huang, W., Liu, Y., Linghu, Q., ... & Hoex, B. (2024). Creation of a structured solar cell material dataset and performance prediction using large language models. Patterns, 5(5).

Maintenance is vital to the smooth operation and safety of any industrial plant. Various maintenance strategies can be employed, often in combination, depending on the industrial sector and the plant's specific operating environment. Common approaches in the European industrial context include corrective, preventive, opportunistic, condition-based, and predictive maintenance (Bevilacqua and Braglia, 2000).

Predictive maintenance is a relatively new concept for used oil refineries, offering the potential to enhance traditional condition-based methods by using machine learning. While scientific literature has applied predictive maintenance based on Gaussian Processes to thermodeasphalting columns with good results, showing reduced runtime in suboptimal regions (Negri et al., 2024), there is room for further improvement through Hybrid Machine Learning (HML).

In this work, an equation-based model is developed to describe the pressure differential (ΔP) along the column, adapting literature models for structured packing (Rocha et al., 1993) to account for fouling through a time-dependent parameter. Given the difficulty in modeling fouling growth due to the changing composition of used oil and the stochastic nature of the phenomenon, this parameter is estimated using Gaussian Process Regressions, providing a most probable growth rate estimation, and a 95% bounded confidence interval. The hybrid model effectively captures the exponential rise in ΔP at the end of the run, which data-driven approaches often missed (Negri et al., 2024).

The model is applied to DCS datasets obtained from the Itelyum Regeneration used oil refinery located in Pieve Fissiraga, Lodi, Italy, and predictive maintenance strategies based on it are proposed and evaluated. These strategies significantly reduce suboptimal column runtime while ensuring economically sustainable operations. This approach is innovative, as hybrid machine learning has not been applied to fouling issues in the used oil industry, offering a more robust maintenance tool that adapts to varying feedstock conditions.

References

Bevilacqua, M., Braglia, M., 2000. Analytic hierarchy process applied to maintenance strategy selection. Reliability Engineering and System Safety 70, 71–83. https://doi.org/10.1016/S0951-8320(00)00047-8

Negri, F., Galeazzi, A., Gallo, F., Manenti, F., 2024. Application of a Predictive Maintenance Strategy Based on Machine Learning in a Used Oil Refinery, in: Manenti, F., Reklaitis, G.V. (Eds.), Computer Aided Chemical Engineering, 34 European Symposium on Computer Aided Process Engineering / 15 International Symposium on Process Systems Engineering. Elsevier, pp. 3175–3180. https://doi.org/10.1016/B978-0-443-28824-1.50530-5

Rocha, J.A., Bravo, J.L., Fair, J.R., 1993. Distillation Columns Containing Structured Packings: A Comprehensive Model for Their Performance. 1. Hydraulic Models. Industrial and Engineering Chemistry Research 32, 641–651. https://doi.org/10.1021/ie00016a010