Combining Bayesian optimization and automation to simultaneously optimize reaction conditions and routes.

Reaching optimal reaction conditions is crucial to achieve high yields, minimal by-products, and environmentally sustainable chemical reactions. With the recent rise of artificial intelligence, there has been a shift from traditional Edisonian trial-and-error optimization to data-driven and automated approaches, which offer significant advantages. Here, we showcase the capabilities of an integrated platform; we conducted simultaneous optimizations of four different terminal alkynes and two reaction routes using an automation platform combined with a Bayesian optimization platform. Remarkably, we achieved a conversion rate of over 80% for all four substrates in 23 experiments, covering ca. 0.2% of the combinatorial space. Further analysis allowed us to identify the influence of different reaction parameters on the reaction outcomes, demonstrating the potential for expedited reaction condition optimization and the prospect of more efficient chemical processes in the future.

Fuelling the Digital Chemistry Revolution with Language Models.

The RXN for Chemistry project, initiated by IBM Research Europe - Zurich in 2017, aimed to develop a series of digital assets using machine learning techniques to promote the use of data-driven methodologies in synthetic organic chemistry. This research adopts an innovative concept by treating chemical reaction data as language records, treating the prediction of a synthetic organic chemistry reaction as a translation task between precursor and product languages. Over the years, the IBM Research team has successfully developed language models for various applications including forward reaction prediction, retrosynthesis, reaction classification, atom-mapping, procedure extraction from text, inference of experimental protocols and its use in programming commercial automation hardware to implement an autonomous chemical laboratory. Furthermore, the project has recently incorporated biochemical data in training models for greener and more sustainable chemical reactions. The remarkable ease of constructing prediction models and continually enhancing them through data augmentation with minimal human intervention has led to the widespread adoption of language model technologies, facilitating the digitalization of chemistry in diverse industrial sectors such as pharmaceuticals and chemical manufacturing. This manuscript provides a concise overview of the scientific components that contributed to the prestigious Sandmeyer Award in 2022.

Quality of uncertainty estimates from neural network potential ensembles.

Neural network potentials (NNPs) combine the computational efficiency of classical interatomic potentials with the high accuracy and flexibility of the ab initio methods used to create the training set, but can also result in unphysical predictions when employed outside their training set distribution. Estimating the epistemic uncertainty of a NNP is required in active learning or on-the-fly generation of potentials. Inspired from their use in other machine-learning applications, NNP ensembles have been used for uncertainty prediction in several studies, with the caveat that ensembles do not provide a rigorous Bayesian estimate of the uncertainty. To test whether NNP ensembles provide accurate uncertainty estimates, we train such ensembles in four different case studies and compare the predicted uncertainty with the errors on out-of-distribution validation sets. Our results indicate that NNP ensembles are often overconfident, underestimating the uncertainty of the model, and require to be calibrated for each system and architecture. We also provide evidence that Bayesian NNPs, obtained by sampling the posterior distribution of the model parameters using Monte Carlo techniques, can provide better uncertainty estimates.

Automated extraction of chemical synthesis actions from experimental procedures.

Experimental procedures for chemical synthesis are commonly reported in prose in patents or in the scientific literature. The extraction of the details necessary to reproduce and validate a synthesis in a chemical laboratory is often a tedious task requiring extensive human intervention. We present a method to convert unstructured experimental procedures written in English to structured synthetic steps (action sequences) reflecting all the operations needed to successfully conduct the corresponding chemical reactions. To achieve this, we design a set of synthesis actions with predefined properties and a deep-learning sequence to sequence model based on the transformer architecture to convert experimental procedures to action sequences. The model is pretrained on vast amounts of data generated automatically with a custom rule-based natural language processing approach and refined on manually annotated samples. Predictions on our test set result in a perfect (100%) match of the action sequence for 60.8% of sentences, a 90% match for 71.3% of sentences, and a 75% match for 82.4% of sentences.

Multilevel Ultrafast Flexible Nanoscale Nonvolatile Hybrid Graphene Oxide-Titanium Oxide Memories.

Graphene oxide (GO) resistive memories offer the promise of low-cost environmentally sustainable fabrication, high mechanical flexibility and high optical transparency, making them ideally suited to future flexible and transparent electronics applications. However, the dimensional and temporal scalability of GO memories, i.e., how small they can be made and how fast they can be switched, is an area that has received scant attention. Moreover, a plethora of GO resistive switching characteristics and mechanisms has been reported in the literature, sometimes leading to a confusing and conflicting picture. Consequently, the potential for graphene oxide to deliver high-performance memories operating on nanometer length and nanosecond time scales is currently unknown. Here we address such shortcomings, presenting not only the smallest (50 nm), fastest (sub-5 ns), thinnest (8 nm) GO-based memory devices produced to date, but also demonstrate that our approach provides easily accessible multilevel (4-level, 2-bit per cell) storage capabilities along with excellent endurance and retention performance-all on both rigid and flexible substrates. Via comprehensive experimental characterizations backed-up by detailed atomistic simulations, we also show that the resistive switching mechanism in our Pt/GO/Ti/Pt devices is driven by redox reactions in the interfacial region between the top (Ti) electrode and the GO layer.

Improved coarse-grained model for molecular-dynamics simulations of water nucleation.

We developed a new coarse-grained (CG) model for water to study nucleation of droplets from the vapor phase. The resulting potential has a more flexible functional form and a longer range cutoff compared to other CG potentials available for water. This allowed us to extend the range of applicability of coarse-grained techniques to nucleation phenomena. By improving the description of the interactions between water molecules in the gas phase, we obtained CG model that gives similar results than the all-atom (AA) TIP4P model but at a lower computational cost. In this work we present the validation of the potential and its application to the study of nucleation of water droplets from the supersaturated vapor phase via molecular-dynamics simulations. The computed nucleation rates at T = 320 K and 350 K at different supersaturations, ranging from 5 to 15, compare very well with AA TIP4P simulations and show the right dependence on the temperature compared with available experimental data. To help comparison with the experiments, we explored in detail the different ways to control the temperature and the effects on nucleation.

Oxidation state changes and electron flow in enzymatic catalysis and electrocatalysis through Wannier-function analysis.

In catalysis by metalloenzymes and in electrocatalysis by clusters related in structure and composition to the active components of such enzymes transition-metal atoms can play a central role in the catalyzed redox reactions. Changes to their oxidation states (OSs) are critical for understanding the reactions. The OS is a local property and we introduce a new, generally useful local method for determining OSs, their changes, and the associated bonding changes and electron flow. The method is based on computing optimally localized orbitals (OLOs). With this method, we analyze two cases, superoxide reductase (SOR) and a proposed hydrogen-producing model electrocatalyst [FeS(2)]/[FeFe](P), a modification of the active site of the diiron hydrogenase enzymes. Both utilize an under-coordinated Fe site where a one-electron reduction (for SOR) or a two-electron reduction (for [FeFe](P)) of the substrate occurs. We obtain the oxidation states of the Fe atoms and of their critical ligands, the changes of the bonds to those ligands, and the electron flow during the catalytic cycle, thereby demonstrating that OLOs constitute a powerful interpretive tool for unraveling reaction mechanisms by first-principles computations.

Simulation of electrocatalytic hydrogen production by a bioinspired catalyst anchored to a pyrite electrode.

The possibility of using the active site, the [FeFe](H) cluster, of the bacterial di-iron hydrogenases as a catalyst for hydrogen production from water by electro- or photocatalysis is of current scientific and technological interest. We present here a theoretical study of hydrogen production by a modified [FeFe](H) cluster stably linked to a pyrite electrode immersed in acidified water. We employed state-of-the-art electronic-structure and first-principles molecular-dynamics methods. We found that a stable sulfur link of the cluster to the surface analogous to that linking the cluster to its enzyme environment cannot be made. However, we have discovered a modification of the cluster which does form a stable, tridentate link to the surface. The pyrite electrode readily produces hydrogen from acidified water when functionalized with the modified cluster, which remains stable throughout the hydrogen production cycle.

Theoretical Design by First Principles Molecular Dynamics of a Bioinspired Electrode-Catalyst System for Electrocatalytic Hydrogen Production from Acidified Water.

Bacterial di-iron hydrogenases produce hydrogen efficiently from water. Accordingly, we have studied by first-principles molecular-dynamics simulations (FPMD) electrocatalytic hydrogen production from acidified water by their common active site, the [FeFe]H cluster, extracted from the enzyme and linked directly to the (100) surface of a pyrite electrode. We found that the cluster could not be attached stably to the surface via a thiol link analogous to that which attaches it to the rest of the enzyme, despite the similarity of the (100) pyrite surface to the Fe4S4 cubane to which it is linked in the enzyme. We report here a systematic sequence of modifications of the structure and composition of the cluster devised to maintain the structural stability of the pyrite/cluster complex in water throughout its hydrogen production cycle, an example of the molecular design of a complex system by FPMD.

Hydrogen production by the naked active site of the di-iron hydrogenases in water.

We explored the reactivity of the active center of the [FeFe]-hydrogenases detached from the enzyme and immersed in acidified water by first-principles Car-Parrinello molecular-dynamics simulations. We focused on the identification of the structures that are stable and metastable in acidified water and on their activity for hydrogen production. Our calculations revealed that the naked active center could be an efficient catalyst provided that electrons are transferred to the cluster. We found that both bridging and terminal isomers are present at equilibrium and that the bridging configuration is essential for efficient hydrogen production. The formation of the hydrogen molecule occurs via sequential protonations of the distal iron and of the N-atom of the S-CH(2)-NH-CH(2)-S chelating group. H(2) desorption does not involve a significant energy barrier, making the process very efficient at room temperature. We established that the bottleneck in the reaction is the direct proton transfer from water to the vacant site of the distal iron. Moreover, we found that even if the terminal isomer is present at the equilibrium, its strong local hydrophobicity prevents poisoning of the cluster.

Mechanism of H2 production by the [FeFe]H subcluster of di-iron hydrogenases: implications for abiotic catalysts.

To explore the possibility that the active center of the di-iron hydrogenases, the [FeFe] H subcluster, can serve by itself as an efficient hydrogen-producing catalyst, we perform comprehensive calculations of the catalytic properties of the subcluster in vacuo using first principles density functional theory. For completeness, we examine all nine possible geometrical isomers of the Fe(II)Fe(I) active-ready state and report in detail on the relevant ones that lead to the production of H 2. These calculations, carried out at the generalized gradient approximation level, indicate that the most efficient catalytic site in the isolated [FeFe] H subcluster is the Fe d center distal (d) to the [4Fe-4S] H cluster; the other iron center site, the proximal Fe p, also considered in this study, has much higher energy barriers. The pathways with the most favorable kinetics (lowest energy barrier to reaction) proceed along configurations with a CO ligand in a bridging position. The most favorable of these CO-bridging pathways start from isomers where the distal CN (-) ligand is in up position, the vacancy V in down position, and the remaining distal CO is either cis or trans with respect to the proximal CO. These isomers, not observed in the available enzyme X-ray structures, are only marginally less stable than the most stable nonbridging Fe d-CO-terminal isomer. Our calculations indicate that this CO-bridging CN-up isomer has a small barrier to production of H 2 that is compatible with the observed rate for the enzyme. These results suggest that catalysis of H 2 production could proceed on this stereochemically modified [FeFe] H subcluster alone, thus offering a promising target for functional bioinspired catalyst design.

Ab initio simulation of the grafting of phenylacetylene on hydrogenated surfaces of crystalline silicon catalyzed by a Lewis acid.

Car-Parrinello simulations have been carried out to identify the grafting mechanism of phenylacetylene, a prototypical alkyne, on the hydrogenated surfaces of crystalline silicon, catalyzed by a Lewis acid (AlCl3). To this purpose, we have made use of a new technique, metadynamics, devised recently to deal with complex chemical reactions in first principles simulations. The reaction mechanism, leading to a styrenyl-terminated surface, turns out to be equivalent to the corresponding gas-phase hydrosilylation reaction by silanes that we have identified in a previous work. The activation energies for the surface reactions (0.43, 0.42, 0.35 eV, for H-Si(111), H-Si(100)2 x 1, and H-Si(100)1 x 1, respectively) are very close to that of the corresponding gas-phase reaction (0.37 eV). The estimated activation free energy at room temperature is sufficiently low for the grafting reaction to be viable at normal conditions and at low coverage on the crystalline silicon surfaces, as already well documented to occur on the surface of porous silicon. However, the conformation of the transition state shadows a large area of the surface, which might contribute to making the grafting process self-limiting.

A QUICKSTEP-based quantum mechanics/molecular mechanics approach for silica.

Quantum mechanics/molecular mechanics (QM/MM) approaches are currently used to describe several properties of silica-based systems, which are local in nature and require a quantum description of only a small number of atoms around the site of interest, e.g., local chemical reactivity or spectroscopic properties of point defects. We present a QM/MM scheme for silica suitable to be implemented in the general QM/MM framework recently developed for large scale molecular dynamics simulations, within the QUICKSTEP approach to the description of the quantum region. Our scheme has been validated by computing the structural and dynamical properties of an oxygen vacancy in alpha-quartz, a prototypical defect in silica. We have found that good convergence in the Si-Si bond length and formation energy is achieved by using a quantum cluster of only eight atoms in size. We check the suitability of the method for molecular dynamics and evaluate the Si-Si bond frequency from the velocity-velocity correlation function.