EAGER: ADAPT: Machine Learning for the Analysis of Novel Zero-field Nuclear Magnetic Resonance Spectroscopic Data

在化学系和多学科活动办公室(OMA)化学理论、模型和计算方法(CTMC)计划的支持下，加州大学伯克利分校(UCB)的阿肖克·阿霍伊和芝加哥大学(U Of C)的埃里克·乔纳斯(Eric Jonas)正在努力开发新的人工智能(AI)方法，用于解释零到超低场(ZULF)核磁共振(NMR)谱。核磁共振波谱是确定未知分子结构的重要和广泛应用的工具，但传统的核磁共振系统工作在强磁场下，设备非常大且昂贵，因此许多研究人员无法获得。传统的核磁共振系统也存在吞吐量非常低的问题。由AJoy实验室率先开发的新型紧凑型ZULF-核磁共振仪器有可能使核磁共振谱更实惠、更广泛地获得和更高的吞吐量。然而，ZULF-核磁共振波谱数据非常复杂，很难解释。在这个项目中，Jonas实验室将建立在他们在解释化学光谱的人工智能技术方面的最新进展，而AJoy实验室将在已知的受控条件下收集多样化的ZULF-核磁共振数据集，用于训练人工智能模型。最后，乔纳斯实验室团队将结合这些技术和数据来产生人工智能模型，该模型可以根据ZULF-核磁共振谱快速确定分子结构。预计这些模型可以扩大规模，在没有人工干预的情况下同时对数十个样本进行自动分析，使“机器人”实验室能够自主发现新的分子物质。阿霍伊和乔纳斯分别在UCB和加州大学的研究小组将合作解决零到超低场(ZULF)核磁共振光谱的光谱到结构的问题。ZULF-核磁共振系统省略了传统高场核磁共振系统中使用的大型、昂贵、高度均匀的超导磁体。这意味着ZULF-核磁共振主要测量核间耦合(J耦合)。由此产生的光谱非常复杂，很难解释。通过将从ZULF-核磁共振光谱确定分子结构视为逆问题，Jonas实验室将首先利用图形神经网络的新发展来创建一个正向模型，该模型可以快速计算给定分子结构的可能光谱。为了为这一前向模型提供训练数据，阿霍伊实验室将获得一组小分子(最多32个原子)的数百个新的实验ZULF-核磁共振谱，而乔纳斯小组将使用从头计算方法模拟数千个其他分子的谱。然后，Jonas团队将使用这个快进模型来模拟数百万分子的ZULF-核磁共振谱，然后使用这些谱分两个阶段训练反向模型：首先，Jonas团队将创建一个模型，在给定观测到的ZULF-NMR谱的情况下，计算自旋系统参数的后验分布；第二，他们将创建一个模型，使用这些自旋系统参数的后验分布通过深度模仿学习来估计分子结构。阿霍伊和乔纳斯的研究小组将通过额外的新实验光谱广泛地验证这一方法。概述的方法有可能在自主实验室中实现自动ZULF-核磁共振结构确定，并且ZULF-核磁共振仪器的低成本可能使结构确定和核磁共振的其他应用可用于更广泛的实验室和新的使用案例。该奖项反映了NSF的法定使命，并通过使用基金会的智力优势和更广泛的影响审查标准进行评估，被认为值得支持。

With support from the Chemical Theory, Models, and Computational Methods (CTMC) program in the Division of Chemistry and the Office of Multidisciplinary Activities (OMA), Ashok Ajoy of the University of California, Berkeley (UCB) and Eric Jonas of the University of Chicago (U of C) are working to develop new artificial intelligence (AI) methods for interpretation of zero-to-ultralow-field (ZULF) nuclear magnetic resonance (NMR) spectra. NMR spectroscopy is a vital and widely applicable tool for determining the structure of unknown molecules, but traditional NMR systems, which operate at high magnetic fields, are very large and expensive pieces of equipment and consequently are inaccessible for many researchers. Traditional NMR systems also suffer from very low throughput. New compact ZULF-NMR instruments pioneered by Ajoy Lab have the potential to make NMR spectroscopy much more affordable, widely accessible, and high in throughput. However, ZULF-NMR spectral data is very complicated and difficult to interpret. In this project, Jonas Lab will build upon their recent advances in AI techniques for interpretation of chemical spectra and Ajoy Lab will gather a diverse ZULF-NMR dataset under known controlled conditions for training of AI models. Finally, the Jonas Lab team will combine these techniques and data to produce AI models that rapidly determine molecular structure from ZULF-NMR spectra. It is anticipated that these models can be scaled up for automated analysis of dozens of samples simultaneously without human intervention, enabling "robotic" laboratories to autonomously discover novel molecular substances. The Ajoy and Jonas research groups at UCB and the U of C, respectively, will collaborate to solve the spectrum-to-structure problem for zero-to-ultralow-field (ZULF) NMR spectroscopy. ZULF-NMR systems omit the large, expensive, highly homogeneous superconducting magnets used in traditional high-field (HF) NMR systems. This means that ZULF-NMR mainly measures inter-nuclear couplings (J-couplings). The resulting spectra are very complex and difficult to interpret. By treating determination of molecular structure from a ZULF-NMR spectrum as an inverse problem, the Jonas lab will first leverage new developments in graph neural networks to create a forward model that rapidly computes the probable spectrum for a given molecular structure. To produce training data for this forward model, the Ajoy lab will acquire hundreds of new experimental ZULF-NMR spectra for a set of small molecules (up to 32 atoms), and the Jonas group will simulate the spectra for thousands of other molecules using ab initio methods. The Jonas group will then use this "fast forward model" to simulate ZULF-NMR spectra for millions of molecules, and then use these spectra to train the inverse models in two phases: first, the Jonas team will create a model to compute the posterior distribution over spin system parameters given an observed ZULF-NMR spectrum; second, they will create a model that uses those posterior distributions of spin system parameters to estimate molecular structure via deep imitation learning. The Ajoy and Jonas research groups will extensively validate this approach with additional new experimental spectra. The outlined approach has the potential to enable automated ZULF-NMR structure determination in autonomous laboratories, and the low cost of ZULF-NMR instruments may make structure determination and other applications of NMR available to a much broader array of laboratories and novel use cases.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.