In heterogeneous catalysis, calculations necessitate intricate computational operations owing to the variability in reaction pathways and options. Now, due to an ingenious fusion of programming and machine learning, researchers have realized a substantial boost in simulation speed, improving the energy efficiency of what is a resource-intensive procedure. The findings, documented for reactions converting carbon dioxide into fuels, could rapidly adapt to other industrially significant reactions, such as depolymerisation and biomass valorisation.
‘We’ve unlocked an understanding that manual simulations could not provide, potentially hastening discovery by several magnitudes,’ states Núria López from the Institute of Chemical Research of Catalonia, who spearheaded the study. This framework aids in predicting selectivity and reactivity in catalysis, especially in the generation of long-chain hydrocarbons from syngas – commonly referred to as Fischer–Tropsch. ‘Previously, computational calculations were hampered by laborious manual tracking of intermediates … along with numerous prolonged processes as alternative reaction pathways emerged,’ she continues. In addition to speed, the new software forecasts properties, or ‘observables’, such as selectivity, reaction velocities, and yield, as clarifies López. ‘This is directly comparable to experimental results, underscoring the programme’s potential,’ she remarks.
In homogeneous and enzymatic catalysis, the active site is typically limited to a small collection of atoms, elucidates materials science and simulation specialist Anastassia Alexandrova from the University of California, Los Angeles, US. ‘In heterogeneous catalysis, the surface is expansive and often intricate, offering a vast array of potential binding sites for reagents,’ she points out. Furthermore, the catalyst’s surface is surprisingly dynamic. As the reaction advances, it undergoes a reconstruction phase ‘under the influence of the reactants … generating numerous different microenvironments for each active site.’ Scanning for possible pathways and networks is ‘insurmountable regarding the necessary time and computational resources, and also prone to errors’ if done manually, observes Alexandrova. ‘This paper marks a significant advancement in the right direction, [investigating] the reaction … more swiftly, leveraging machine learning.’
What are AI, machine learning and neural networks?
Artificial intelligence (AI) is a broad term frequently misused to encompass a range of connected but simpler processes. AI signifies the capability of machines and computer programs to execute tasks that typically only humans could perform, such as reasoning, responding to feedback, and making decisions. Generative AI is a newer variant of AI that analyzes and identifies patterns in training datasets to create original text, images, and videos in response to user requests. ChatGPT, Microsoft Copilot, Google Gemini, and more recently X’s Grok are examples of chatbots that utilize generative AI.
Neural networks are a connected network of artificial neurons, similar to biological brains, that recognize, evaluate, and learn from statistical patterns within data. Machine learning is a subdivision of AI that enables machines to learn from datasets and predict outcomes based on new data, without explicit instructions from programmers. Machine learning models enhance their performance as they accumulate more data. Deep learning is an advanced form of machine learning that utilizes neural networks with multiple layers to assess complex data from very large datasets. Applications of deep learning encompass speech recognition, image creation, and translation. Large language models or LLMs are a type of deep learning trained on substantial amounts of data to comprehend and produce language. LLMs learn text patterns by forecasting the subsequent word in a sequence, and these models are now capable of crafting prose, analyzing internet text, and engaging in dialogues with users.
‘This revolutionary framework automatically charts and examines enormous, complex chemical reaction networks that were previously too vast or troublesome to manage manually,’ articulates Ritesh Kumar, an expert in science and artificial intelligence at the University of Chicago, US. The system ‘quickly and accurately forecasts’ the reactivity of surface catalysts, ‘without a scientist needing to guess every possible step’, he notes. ‘It truly substitutes guesswork with intelligent automation, swiftly estimating the energy and pace of thousands of steps.’
Whereas conventional density functional theory (DFT) programs forecast up to 500 steps in 100 processing hours, this new solution accelerates the search by significant orders of magnitude – modeling 370,000 potential pathways within a comparable timeframe. ‘The velocity is remarkable – it identifies critical reactions in a fraction of the time and expense,’ asserts Kumar, who will soon commence a role at TCG Crest, India. Furthermore, ‘without the vast energy consumption typically demanded by supercomputers’. In addition to the sustainability advantages, the automated algorithms could enable scientists to allocate specific – and slower – computing resources for crucial calculations only, explains Kumar. ‘In processes like Fischer–Tropsch, the number of potential pathways explodes into the hundreds of thousands … with traditional approaches, it could take centuries to compute,’ he notes. Now, neural networks could autonomously uncover reaction pathways and accelerate the examination of complex catalytic processes.