The reproducibility of chemical reactions, especially photochemical ones, poses a significant hurdle in academic and industrial settings. Obtaining consistent results from published protocols is often difficult, hindering both research and development. Synthesizing organic molecules under photochemical conditions typically requires extensive optimization, highly skilled personnel, and lengthy reaction times, making the process costly and inefficient. This challenge is further amplified by the complexity of photochemical reactions, where numerous parameters such as light intensity, temperature, pressure, reaction time, quantum efficiency, and reagent concentrations must be carefully controlled to achieve desired outcomes. The manual optimization process is often time-consuming, relying heavily on intuition and experience. Existing databases of chemical synthesis protocols, while useful, often lack the detail necessary for successful replication. This paper addresses this challenge by introducing DigiChemTree, a novel platform that integrates continuous flow reactions with artificial intelligence (AI) to automate the optimization of light-induced carbene generation reactions. Light-induced reactions of diazo compounds are particularly crucial in organic synthesis and drug discovery, as they allow for novel transformations and exploration of unexplored chemical space. While traditionally, thermal conditions were used, leading to low reactivity and selectivity, the use of transition metal carbenes offered some improvement, but this approach is often limited by cost and impurity generation. Photochemical methods, while offering potential, face challenges including the need for high-energy UV irradiation, expensive filters, side reactions, and specialized equipment. DigiChemTree aims to overcome these limitations by using a closed-loop Bayesian optimization (BO) approach to systematically explore and optimize multiple reaction parameters simultaneously, facilitating efficient and reproducible synthesis of a library of molecules on demand. This platform promises to accelerate chemical synthesis, particularly in areas such as drug discovery and late-stage functionalization of active pharmaceutical ingredients (APIs).

The literature extensively covers light-induced reactions of diazo compounds, highlighting their importance in organic synthesis and drug discovery. However, many studies remain confined to academic settings, with limited industrial applications due to the challenges associated with these reactions. Several papers describe various methods for carbene generation, including thermal methods, transition metal catalysis, and photochemical approaches. However, the complexity of optimizing multiple reaction parameters and the need for specialized equipment have hindered broader application. Recent advances in AI-driven reaction optimization offer a potential solution, with studies demonstrating the use of machine learning to accelerate and improve reaction design. The successful application of AI to other aspects of chemical synthesis, such as route planning and process intensification, further supports the potential of DigiChemTree. The existing literature lacks a unified platform that combines continuous flow chemistry with AI-driven optimization for photochemical carbene generation reactions. DigiChemTree fills this gap by offering an automated, high-throughput, and reproducible method for synthesizing a wide range of molecules.

DigiChemTree is a fully automated platform designed for optimizing and performing light-induced carbene generation reactions. The system consists of several key components: nine liquid delivery pumps for precise reagent control, a PFA micro-flow tubular reactor for reaction execution, a circular blue LED photo-light source for controlled light irradiation, a controlled power supply for adjusting light intensity, affordable com port connectivity for system integration, a 3D-printed solution collector for collecting products, and an in-line IR system for real-time reaction monitoring. The platform operates based on a closed-loop Bayesian optimization (BO) approach. This involves initially defining boundaries for parameters such as flow rate, voltage, current, and reaction time. The BO algorithm then iteratively suggests new experiments, aiming to maximize the reaction yield. The platform uses Python code (with specific code names like ab1, ab2, etc. for different reactions) to control the system and collect data. An in-line IR background analysis is performed before the experiments to determine the signature peak for product identification. The system mixes the reactant solutions at a T-junction, and the mixture then flows through the reactor. Blue LED light of varying intensity is used, and the reaction is conducted under controlled pressure. The collected data, including yield and reaction parameters, are fed back to the BO algorithm, which adjusts the parameters for the subsequent experiment. This closed-loop system continues until the yield is maximized or a predefined termination criterion is met. The optimized parameters are then used for scale-up experiments to synthesize larger quantities of the desired products. The authors describe using the system for various carbene insertion reactions with different nucleophiles (O-H, S-H, N-H, alkenes, alkynes, and acetonitrile). The methodology involved optimizing the reaction conditions for each nucleophile, including adjusting flow rate, residence time, light intensity and pressure to maximize yield. This is done using the Python code, which is given in the supplementary information. Specific codes (AB1-AB7) were developed for each reaction type, and later an integrated code was developed to enable the synthesis of multiple products without human intervention. Different solvents such as EtOAc and DCE were employed depending on the reaction. The platform was used for late-stage functionalization of bioactive molecules as well.

DigiChemTree successfully automated the optimization of several photochemical carbene insertion reactions. The platform consistently achieved high yields (82-99%) and significantly reduced reaction times compared to traditional batch processes. For example, the C-O bond formation reaction reached 99% conversion in a short residence time with a space-time yield of 0.0909 g mL⁻¹ h⁻¹, significantly outperforming the conventional batch process (0.0007 g mL⁻¹h⁻¹). Similarly, impressive results were obtained for C-S, C-N, C-C, and N-C-O bond formations. In each case, the AI-driven optimization identified optimal conditions with increased yield and significantly reduced reaction times. The platform showed adaptability across different reaction types and substrate scopes, demonstrating its broad applicability. The system successfully synthesized a wide range of products, many of which are novel compounds. Out of the 16 synthesized photo carboxylation products, 15 were new. Similarly, 6 out of 8 photo α-thiocarbonyl products and 3 out of 9 cyclopropanated products were new. This highlights the platform's capability in accessing unexplored chemical space. The optimized conditions and the corresponding Python codes (AB1-AB7) were used to produce various molecules, with yields ranging from 75% to 99%. The successful optimization and synthesis of several cross-coupling products were demonstrated, showcasing the platform's versatility in generating diverse molecular structures. The system's ability to perform late-stage functionalization of bioactive molecules further strengthens its potential in drug discovery and development. The overall process was efficient with minimal human intervention, demonstrating the capability of AI-driven automated synthesis for complex chemical transformations. The integrated platform, termed "Innovative DigiChemTree," enabled on-demand synthesis of cross-coupling molecules in excellent yields without any human intervention.

DigiChemTree addresses the critical need for reproducible and efficient chemical synthesis. The AI-driven automation significantly reduces the reliance on highly trained personnel and time-consuming manual optimization, making complex chemical synthesis more accessible. The platform's high yields and reduced reaction times offer significant advantages over traditional methods, particularly for industrial applications. The successful late-stage functionalization of bioactive molecules showcases its potential in medicinal chemistry and drug discovery. The ability to explore unexplored chemical space highlights the potential for discovering novel molecules with desired properties. The development of the integrated platform streamlines the synthesis process and reduces the need for individual optimization for each reaction. The current study demonstrates the feasibility of applying AI-driven automation to photochemical reactions. The platform can be further improved by integrating more advanced AI algorithms, expanding the range of reactions and substrates, and incorporating in-line analysis techniques for better process monitoring and control. Future development should focus on user-friendly software and incorporating broader reaction types to maximize versatility.

DigiChemTree represents a significant advancement in automated chemical synthesis. The platform’s AI-driven optimization and continuous flow operation enable rapid, efficient, and reproducible synthesis of a wide range of molecules. The high yields, short reaction times, and ability to perform late-stage functionalization highlight its potential for accelerating drug discovery and other applications. Future work will focus on developing a user-friendly interface and expanding the platform's capabilities to encompass even more complex chemical reactions.

While DigiChemTree demonstrates significant advancements, some limitations exist. The current platform is optimized for specific reaction types and may require modifications to accommodate different reaction chemistries. The system's reliance on specific equipment and software might limit accessibility for researchers with limited resources. Further research is needed to explore the platform's applicability to a broader range of reactions and substrates. The long-term stability and robustness of the system need further evaluation for industrial scale-up. Additionally, a user-friendly interface is needed to make the system more widely accessible.