Alexei A. Lapkin

Conspectus: Today we are experiencing an unprecedentedly rapid transformation of science, driven by advances in computing and artificial intelligence (AI). At the same time, there are growing cumulative pressures of climate change impacts that not only affect economies but are creating significant societal challenges. The confluence of technology push and environmental–socioeconomic pressures is promoting the development of sustainable chemistry. The link between the transition to sustainability in chemistry and the digital transformation of R&D and manufacturing has been the long-standing focus in our research (Fantke, P.; et al. Chem 2021, 7, 2866 doi: 10.1016/j.chempr.2021.09.012). Sustainability is not a feature of a single chemical product or a process. It is incorrect to talk about “sustainable polyethylene”, as such narrow framing will miss many issues in the more holistic system that not only includes the cradle-to-grave life cycle emission impacts of polyethylene but also considers all aspects of equity, impact on biodiversity, resilience of supply chains, etc. Sustainable chemistry makes sense only as a complex system. An analysis of complex systems requires large amounts of data and the use of appropriate analytical tools, such as advanced machine learning techniques. The chemical industry has been an early adopter of digital technologies, motivated by the inherent complexity of chemical systems and the need for efficiency to stay economically competitive. As early as the 1970s, cheminformatics tools were introduced to support structure–property analysis. In the 1980s, we saw the broad adoption of process modeling and simulation tools for design, scale-up, and optimization of chemical processes, paralleled with developments in the field of process system engineering and control. Today’s digital transition is much deeper; everything from scientific discovery (molecules, materials, and phenomena) to chemical process development, analytical methods and process analytical technologies, and plant operation is being reshaped. This deep transformation includes the development of a radically different approach to data in chemistry. Advances in self-driving laboratories, agentic AI workflows, and digital factories require access to structured, semantically rich data. In this Account, we focus on three key areas of digitalization and AI technologies advancing chemistry and chemical process development toward net-zero targets. First, we discuss how data mining tools are used to analyze the existing value chains and identify opportunities to integrate net-zero feedstocks. When new molecules are introduced, data-driven route planning can be employed to design alternative synthetic pathways, which then need to be rigorously evaluated. Second, we focus on how to use AI to accelerate discoveries in chemistry, including the innovative design of catalysts, rational selection of solvents, and identification of optimal reaction conditions. By leveraging statistical modeling and data-driven optimization, these traditionally time-consuming tasks can now be performed much faster. Lastly, we discuss how to bring these innovations from the laboratory to manufacturing, which also requires advances in process technologies. We show that by leveraging AI agents, process modeling, and knowledge graphs, there is a potential to achieve automated process design and life cycle assessment, removing traditional choke points in scaling technologies and providing timely support for decision-making in early-stage development. In the end, we discuss the current challenges and future perspectives.

To date the environmental impacts of production of most chemicals are unknown, hindering sustainable process evaluation and design. Life cycle assessment (LCA) is the most comprehensive method for estimating these environmental impacts but existing LCA databases cover a small fraction of the industrial chemicals. Environmental impact prediction models promise to bridge this gap; several models were published, with the apparent limitations of limited accuracy and applicability. The very small number of available LCA datasets for industrial chemicals is the major issue for developing environmental impact prediction models; this limits the domain of applicability and accuracy of the models. Here we show that a machine learning approach using transfer learning partially addresses these limitations. We found supervised pretraining on molecular price data increases model accuracy for three different deep learning model architectures, for two out of three different environmental impact categories. Taking the mean for model architectures, across all impacts, transfer learning increases R ² by 0.11, and reduces mean percentage absolute error by 2.5%. Our most accurate model for carbon footprint prediction achieves an R ² of 0.62, similar to existing literature benchmarks, and an overall prediction error of 22% for the organic precursors for six active pharmaceutical ingredients.

We formulate the discovery of aqueous redox flow battery catholytes as a multi-objective optimization problem and introduce a virtual screening framework that simultaneously considers redox potential, chemical stability, and synthetic accessibility. Applied to a library of 182 TEMPO derivatives, the workflow identified a diverse set of high-ranking candidates, including the previously unexplored methyl carbamate-TEMPO (MC-TEMPO) and diethylamino-TEMPO (DEA-TEMPO). Furthermore, the screening revealed a moderate correlation between redox potential and predicted degradation resistance, indicating an intrinsic trade-off between these properties. For established TEMPO derivatives, the framework reproduced relative performance trends reported within independently studied series. Across the broader library, the highest-ranked candidates reflected a non-intuitive balance of electrochemical performance, stability, and synthetic accessibility rather than optimization of redox potential alone. MC-TEMPO and DEA-TEMPO were subsequently synthesized and experimentally characterized, providing a prospective test of the computational predictions. Cyclic voltammetry validated the redox potential predictions, with measured values closely matching the calculated values. During 130 h potentiostatic holds under sustained oxidizing conditions, MC-TEMPO showed no measurable electrochemical or molecular degradation, whereas DEA-TEMPO underwent a substituent-specific degradation pathway not captured by the initial screening model. Together, these results show that multi-objective catholyte discovery can prioritize molecules that would be overlooked by redox-potential optimization alone and provides a framework for the computationally guided design of organic redox flow battery catholytes.

Accurate prediction of molecular properties is central to advancing chemistry, materials science, and drug discovery. Machine learning on molecular graphs depends critically on representations that capture the topology and structure of molecules. Here we propose the dual graph transformer (DGT), a self-attention architecture that jointly models atom and bond graphs to achieve comprehensive molecular encodings. DGT fuses atom and bond features, graph topology and structure, and stereogeometric information within its self-attention module for an effective molecular representation. We benchmark DGT across a range of datasets for molecular property prediction, showing that it considerably outperforms the current state of the art. DGT demonstrates performance contributions from its dual graph representation, relative positional and structural encodings, and stereogeometric information incorporation while also offering interpretability at the molecular structural level. We envision DGT advancing molecular machine learning by improving both the prediction accuracy and interpretability of molecular properties.

A digital twin, which virtually replicates a real system and fuses data, models and domain knowledge, is a key technology for accelerating chemical process development and addressing sustainability challenges. Despite its potential, one critical challenge lies in the lack of a systematic approach to integrate data, domain knowledge and predictive models to contextualize and represent chemical processes effectively. Here we propose a knowledge graph framework associated with autonomous functional agents to support the development of digital twins for chemical processes, enabling the seamless incorporation of chemical databases, artificial intelligence models and large language models. Ontologies are developed for physical models of chemical processes, allowing scalable model construction and calibration. We demonstrate the framework with practical case studies focusing on bottom-up model assembly, top-down model search and model-based reaction optimization. The framework presents an approach to manage models as a depository of chemical process knowledge, providing a foundation of digital twin technology for future chemical process development and manufacturing.

Designing experimental campaigns to make efficient use of available time and resources is a critical task in any researchers day. For complex chemical engineering problems, model-based design of experiments (MBDOE) is a reliable tool to sequentially conduct optimal experiments. However, since this paradigm is built upon frequentist methods, uncertainty of data is oftentimes not considered in it despite being critical. On the other hand, Bayesian statistics methods allow for uncertainty modeling throughout the experimental design workflow. Here, a Bayesian design of experiments (bDOE) workflow is introduced leveraging Bayesian optimization (BO) in conjunction with three information-theory based acquisition functions to design optimal sequential experimental plans. It is demonstrated that the experimental budget can be reduced by approximately 60 % when parameterizing kinetic models using this methodology. Model discrimination can be achieved within very few (approximately four) specifically designed experiments. Overall, the proposed bDOE workflow lends itself to problems where the model structure is known but experiments are time-consuming, expensive or limited in number.

Accurate prediction of molecular properties is central to advancing chemistry, materials science, and drug discovery. Attention-based transformers have recently emerged as an effective infrastructure for learning on molecular graphs, yet their full potential depends on representations that capture the rich topology and structure of molecules. To this end, we propose the dual graph transformer (DGT), a self-attention architecture that jointly models atom and bond graphs to achieve comprehensive molecular representation. DGT facilitates the fusion of atom and bond features, graph topology and structure, and three-dimensional information within the self-attention module to learn an effective molecular representation. We benchmark DGT across a range of datasets for molecular property prediction, showing that it considerably outperforms the current state of the art. DGT demonstrates performance contributions from its dual graph representation, relative positional and structural encodings, and spatial information incorporation, while also offering interpretability at the molecular structural level. We envision DGT advancing molecular machine learning by improving both the prediction accuracy and interpretability of molecular properties.