(248l) Data-Driven Design and Identification of Active Site Motifs in Amorphous Porous Polymer Catalysts Using Molecular Simulation and Reinforcement Learning

Amorphous porous organic polymers (POPs) offer significant design versatility for specific catalytic applications due to their flexible and shape-selective nanopore environments compared to zeolites and Metal-Organic Frameworks (MOFs). In particular, POPs functionalized with acid sites can catalyze esterification reactions¹ and selective organic conversions, including the formation of cyclic ester carbonates from CO₂². To effectively evaluate the catalytic performance of a POP computationally, its structure must be simulated, active sites motifs identified, and after which Density Functional Theory (DFT) can be used to calculate the adsorption or activation energies. The intricate and amorphous nature of POPs complicates this conventional computational catalysis workflow, resulting in challenges for feature engineering and high computational costs for high-throughput screening. We introduce a strategy comprising two cooperative methodologies to address these limitations. The first part of our strategy is applying a reinforcement learning (RL) algorithm to streamline the discovery and optimization of potential catalysts within POP networks. This approach distinguishes itself from conventional machine learning approaches that typically require extensive labeled databases and may not effectively handle the dynamism and complexity in catalyst design. Our approach leverages RL to efficiently navigate the design of POPs to generate polymers with desired catalytic properties. We generated structures by simulating the polymerization process using Polymatic³, in which monomers were packed and bonded using molecular dynamics to create an amorphous polymer structure. We then used RL to modify polymerization conditions and assess the obtained pore environments. The second part of our strategy uses image recognition and clustering tools to identify amorphous active site motifs based on their local environments generate surrogate models that are tractable with Density Functional Theory (DFT). Our ML model integrates these two components for high-throughput structure generation and screening. By prioritizing kinetically relevant regions within the polymer networks, our AI model improves the efficiency of identifying potential active sites in POPs.

Our RL environment training results demonstrate the capability to improve the adsorption properties of polymers and to target specific catalytic properties as training objectives. In a simulation environment using styrene, divinylbenzene, or porous aromatic frameworks (PAFs) as example initial monomers at different mixing ratios⁴, our model aligns its learning trajectory towards improving desired textural properties, such as average pore size, pore-limiting diameter, and the completion rate of polymerization, by setting these textural features as goals. For identification of active site motifs, our AI model incorporates the screening workflow that combines Density Functional Theory (DFT) and Monte Carlo simulations, streamlined for efficiently computation of applications such as CO2 cyclization and ammonia decomposition. Tools such as PoreBlazer are implemented in our model for analyzing structural properties, offering a comprehensive assessment of catalytic capabilities. This screening methodology enables our model to effectively improve and update the training policies. The results show the improvement of porous structure and catalytic capabilities of polymers simulated in an RL environment by the ML workflow, highlighting the efficacy of this approach in enhancing polymer design for catalysis and adsorptive separations.