Designing better catalysts with HPC – DIC, SandboxAQ, and AWS collaborate for record quantum chemistry computation

Catalysts play an important role in accelerating and improving the selectivity of chemical reactions, making them essential for the scalable and sustainable production of advanced materials, such as fine chemicals. Establishing effective procedures for catalysis is therefore a critical step, especially in the industrial field.

Traditionally, catalyst design has posed a substantial challenge, relying heavily on the empirical knowledge and manual experimentation of experienced chemists, often involving extensive trial and error. One possible way to streamline catalyst design is by leveraging computational methods, an approach widely adopted in the field of drug design. However, accurately guiding catalyst design with computational chemistry calculations is difficult due to the complexity involved in modeling these systems, primarily because of the prevalence of transition metal elements and the resulting multitude of electron configurations.

In this post, we’ll talk about a new approach to this challenge. It involves a collaboration with a global chemical manufacturer, DIC, and SandboxAQ, a SaaS solutions company at the convergence of artificial intelligence and quantum technologies. With help from AWS, these companies are working to accurately solve the electronic structure of organometallic compounds at the industrially-relevant molecular scale for potential catalyst candidates.

The challenge of computational catalyst design

The intricacy of computational catalyst design arises from the complicated electronic structure of organometallic compounds. This impedes the effectiveness of commonly used approximate quantum chemistry methods in accurately capturing the catalytic properties of these chemical systems (both density functional theory and coupled cluster are widely used) .

On the other hand, obtaining exact solutions for chemical properties via the full configuration interaction (FCI) method is intractable due to its exponential scaling with system size (see Figure 1).

In addition, efforts to study organometallic catalysts often involve a simplification of the molecular structure to stay within computational limitations. However, the completeness of molecular models plays a crucial role in the catalyst design process, which is based on accurate predictions of a system’s structure and behavior, as it is essential to align computational results with experimental findings.

Using simplified molecular structures in the analysis can fail to accurately represent real catalysts by excluding significant electronic and steric interactions, which undermines the integrity of catalyst design. Employing structures that mimic the size of actual catalysts in our study, rather than simplified models, necessitates substantial computational resources due to the increased complexity and scale of the calculations involved. This approach, while enhancing the accuracy of our models in reflecting real-world catalysts, markedly elevates the computational demands.

Figure 1. FCI scales exponentially with system size. Even the smallest molecule presented here (C2H4) with a modest basis set (6-31G*) is difficult to compute (10^12 configurations), while the smallest organometallic catalyst considered in this study (Complex 1a) is far past computational feasibility (10^139 configurations).

A novel high throughput cloud-native platform

QEMIST Cloud is a computational chemistry platform developed by SandboxAQ. It’s a high throughput cloud-native software as a service (SaaS) solution built on AWS infrastructure (Figure 2).

The container-based cloud HPC architecture is based on services like Amazon Elastic Kubernetes Services (Amazon EKS), Karpenter, and Amazon Aurora databases with Amazon RDS Proxy. QEMIST Cloud enables us to perform highly accurate and scalable quantum chemistry calculations by using the incremental full configuration interaction (iFCI) method. iFCI is a novel, highly scalable approximation of the FCI method, allowing convergence toward the FCI solution. This algorithm decomposes a molecule into sub-problems that can be solved independently from each other, making it extremely suitable for massive parallelization.

Recently, iFCI on QEMIST Cloud was used to simulate the bond breaking of PFAS molecules using over 1 million vCPUs on AWS Cloud, pushing the boundaries of what’s possible for electronic structure computations in the cloud.

Figure 2. High-level architecture of the HPC cluster and iFCI application in the AWS environment (top) incorporated into QEMIST Cloud (bottom), with one of the catalysts studied shown on the dashboard.

Record-breaking results

We applied iFCI in QEMIST Cloud on AWS to a series of organometallic catalyst candidates containing nickel and investigated their energy differences between singlet and triplet spin states.

These molecules, or complexes, consist of a central metal atom bonded to three ligands (bonding groups) of varying size and composition, and serve as examples of novel non-precious metal catalysts (Figure 3).

Solving for any of these catalysts is intractable if computed with FCI, where the largest (Complex 3) would require solving 10²⁴² configurations (using the 6-31G* basis set). However, even the largest of these complexes becomes tractable with iFCI.

For example, we scaled QEMIST Cloud to 2,200 workers on c6i.4xlarge instances from Amazon Elastic Compute Cloud (Amazon EC2) to compute Complex 3’s singlet state, which involved computing roughly 29,000 sub-problems with a cumulative run time of almost six months. This took just over 5 hours, making it the largest organometallic catalyst system to be calculated at this level of accuracy, to date.

Figure 3. Ni complexes in this study.

Table 1. Simulated molecules with computational results.

The energy difference between the singlet and triplet spin states -the spin gap – is one of the most important metrics for understanding the catalytic activity of the candidate metal complexes (Table 1).

During our analysis, we observed that iFCI suggests that Ni complexes with more nitrogen atoms coordinated to the metal center (Figure 1b, 2b, and 3) showed higher catalytic activity compared to those with two oxygen atoms replacing two nitrogen atoms (Figures 1a and 2a), which is consistent with experimental observations for other types of catalysts.

Conventional DFT was not able to reproduce this trend, exhibiting different catalytic behavior from what is seen in iFCI and experimental results. The present results demonstrate the near-exact accuracy and scalability that are essential for computational catalyst design, which are now accessible via QEMIST Cloud on AWS.

Conclusion

In computational catalyst design, the most important tasks are: to accurately predict the reactivity of the catalyst candidates; and determine how a small change in structure, like the replacement of one element with another, makes an impact on its catalytic reactivity.

However, such accurate computation has been limited to smaller model systems due to the computational complexity we described in this post. Our results indicate that the highly accurate and scalable QEMIST Cloud on AWS breaks this limitation and makes it possible to perform accurate computational catalyst design at the industrially-relevant molecular scale.

We plan to extend this work by taking advantage of QEMIST Cloud’s compatibility with Tangelo and Amazon Braket (the quantum computing service of AWS) to explore the potential of quantum computing to accurately model the electronic structure of catalysts.

Through this collaboration, we aim to precisely control chemical reactions essential for DIC’s product development with a primary goal of achieving high catalytic efficiency using abundant, non-precious, and nontoxic metals – for a more sustainable future.