Abstract
The discovery of crystalline reticular materials remains largely trial-and-error despite their societal importance. We introduce our algorithmic iterative reticular synthesis (AIRES) cycle, which integrates automated synthesis, image recognition, single-crystal X-ray diffraction and, crucially, customized algorithmic decision-making, to maximize distinct crystal discoveries rather than optimizing single targets. Demonstrated on zeolitic imidazolate frameworks (ZIFs), AIRES achieves twice the discovery rate of random exploration, crystallizing 10 new linkers into diverse ZIF topologies and expanding the single-linker Zn-ZIF library by one-third. By transforming reticular synthesis from an empirical process to a systematic exploration, AIRES provides a scalable and efficient blueprint for accelerating materials discovery.
Introduction
The discovery of new crystalline reticular materials, such as metal–organic frameworks (MOFs) and zeolitic imidazolate frameworks (ZIFs), is a cornerstone of modern materials science. However, the process remains largely governed by trial-and-error. While building blocks can be designed rationally, predicting the specific conditions—temperature, concentration, and solvent—required for crystallization remains an empirical "black box." Traditional machine learning (ML) in chemistry often focuses on optimizing the properties of known materials. In contrast, the goal of discovery is to maximize the number of distinct new structures found within a limited experimental budget.
The AIRES Cycle
To address this bottleneck, we developed the Algorithmic Iterative Reticular Synthesis (AIRES) cycle. This integrated loop replaces human intuition with a systematic, robotic search strategy.
The cycle consists of four key pillars:
- Automated Synthesis: An Opentrons OT-2 robot prepares reaction mixtures in high-throughput 96-well plates.
- Computer Vision: An EfficientNetV2-S convolutional neural network analyzes microscope images of reaction products, classifying them as "crystal" or "non-crystal" with a 0.88 recall rate.
- Characterization: Single-crystal X-ray diffraction (SCXRD) provides definitive structural solutions for identified crystals.
- Algorithmic Decision-Making: Utilizing a "greedy" selection strategy, the model (Random Forest or Gaussian Process) predicts crystallization probabilities. Once a linker successfully crystallizes, the algorithm pivots to unexplored linkers to maximize discovery diversity.
Mapping the Crystallization Landscape
The effectiveness of AIRES relies on high-quality data curation. The system was trained on an initial set of 2,688 reactions using 8 starting linkers and an image dataset of over 20,000 labeled images.
By mapping these binary outcomes (success vs. failure), the AIRES model learns to identify patterns across different chemical families. For instance, it can apply insights gained from imidazoles to accelerate the discovery of benzimidazole-based structures.
Key Results and Discovery Efficiency
AIRES demonstrated a significant performance leap over traditional methods. In a validation set of 1,728 reactions across 48 new linkers, AIRES discovered all 10 new linker systems in approximately 700 experiments. In comparison, a random search required 1,400 experiments to achieve the same result, representing a 2x speedup in discovery efficiency.
This systematic exploration expanded the single-linker Zn-ZIF library by 33%, adding 10 new linkers and 11 new crystal structures across 7 distinct topologies (SOD, ANA, DFT, RHO, ykh, crb, and a novel "double fes" motif).
Conclusion and Future Outlook
The AIRES cycle transforms reticular synthesis from an empirical art into a scalable, algorithmic process. While currently limited by linker solubility and certain functional group interferences, future iterations aim to incorporate broader chemical spaces, including multi-linker ZIFs, MOFs, and COFs. By prioritizing discovery over optimization, AIRES provides a blueprint for breaking the "discovery plateau" in materials science.
Summary updated