EAGER: ADAPT: Optimizing Chemical Reaction Networks With AI

西北大学的托德金里奇得到了美国国家科学基金会数学和物理科学理事会、人工智能计划和化学部的一项奖励，以开发和评估用于优化反应扩散化学的人工智能算法。例如，反应和扩散的结合使生命系统能够调节基本过程，例如，信号是如何在大脑中处理和传播的生物系统不是通过电子流过计算机芯片和电线来发挥作用，而是通过分子相互转换（反应）和分子在空间中扩散来发挥作用。然而，不清楚如何最好地混合必要的反应和扩散以实现期望的功能。设计反应扩散化学的一个重要障碍是，单个反应和扩散事件具有一定的随机性，并且计算模拟必须在存在噪声随机波动的情况下优化化学。Gingrich博士和他的研究小组正在寻求计算方法，通过开发新的算法来减轻噪音，该算法利用一种称为张量网络的数学结构来有效地平均噪音。作为另一个例子，可以产生一些化学反应，其充当分子在低浓度和高浓度之间振荡的时钟。 正在开发的方法是人工智能工具，可以识别调节反应扩散化学以调节振荡的策略。这些基于iTensor软件库的技术进步将被公开和自由传播。该研究将与Radboud大学（NL）专门研究人工智能（AI）的Tal Kachman小组合作完成。该项目旨在开发AI算法，该算法通过基于梯度的搜索来识别速率常数以优化目标函数，其中梯度测量改进（例如，在化学振荡器中）由于基元反应速率的微小变化。核心技术挑战是计算这些梯度，这是一个需要精确和有效的数值解决方案的问题，用于非常高维空间中的化学动力学。一种简单的方法是利用确定性的耦合微分方程来求解质量作用动力学，但众所周知，涌现的化学反应网络现象只能通过包含化学动力学随机性质的算法来捕获。像吉莱斯皮算法这样的算法生成化学动力学的随机实现，并通过对许多噪声轨迹进行平均来实现准确性。Gingrich博士和他的研究小组利用Doi-Peliti形式主义与量子动力学的类比，正在开发和分析一种替代方法，该方法使用张量网络来有效地平均所有可能的轨迹。该奖项反映了NSF的法定使命，并通过使用基金会的智力价值和更广泛的影响审查标准进行评估，被认为值得支持。

Todd Gingrich of Northwestern University is supported by an award from the by an award from the NSF Directorate of Mathematical and Physical Sciences, Artificial Intelligence Program, and the Division of Chemistry, to develop and assess AI algorithms for optimizing reaction-diffusion chemistry. As an example, the combination of reaction and diffusion enables living systems to regulate essential processes, e.g., how signals are processed and propagated in the brain. Rather than functioning with electrons flowing through computer chips and wires, biological systems perform functions by molecular interconversion (reactions) and molecular diffusion through space. It is not, however, clear how to best mix the necessary reactions and diffusion to achieve a desired function. A significant barrier to designing reaction-diffusion chemistry is that the individual reactions and diffusion events occur with some randomness, and computational simulations must optimize the chemistry in the presence of noisy stochastic fluctuations. Dr. Gingrich and his research group are pursuing computational approaches to mitigate the noise by developing new algorithms that utilize a mathematical construction called a tensor network, to effectively average over the noise. As another example, some chemical reactions can be generated that act as a clock with a molecule oscillating between low and high concentration. The methods being developed are AI tools that would identify strategies to modulate the reaction-diffusion chemistry to regulate the oscillations. Those technical advances, built upon the iTensor software library, will be openly and freely disseminated. The research will be done in collaboration with the group of Tal Kachman, specializing in artificial intelligence (AI) at Radboud University (NL).This project aims to develop AI algorithms that identify rate constants to optimize an objective function by gradient-based search, where the gradients measure improvements (e.g., in a chemical oscillator) due to small changes in the elementary reaction rates. The core technical challenge is to compute those gradients, a problem that demands accurate and efficient numerical solutions, for chemical kinetics in exceptionally high-dimensional space. A naïve approach utilizes deterministic, coupled differential equations for the mass-action kinetics, but it is well-known that emergent chemical reaction network phenomena are only captured by algorithms that incorporate the stochastic nature of chemical kinetics. Algorithms like the Gillespie algorithm generate stochastic realizations of chemical kinetics and achieve accuracy by averaging over many noisy trajectories. By utilizing the Doi-Peliti formalism’s analogies with quantum dynamics, Dr. Gingrich and his research group are developing and analyzing an alternative method that uses tensor networks to effectively average over all possible trajectories.This award reflects NSF's statutory mission and has been deemed worthy of support through evaluation using the Foundation's intellectual merit and broader impacts review criteria.