电池电解质的大数据研究方法

Discovering new, powerful electrolytes is one of the major bottlenecks for designing next-generation batteries for electric vehicles, phones, laptops and grid-scale energy storage.
发现新型强效电解质是设计电动汽车、手机、笔记本电脑和电网级储能下一代电池的主要瓶颈之一。

The most stable electrolytes are not always the most conductive. The most efficient batteries are not always the most stable. And so on.
最稳定的电解质并不总是导电性最强的。最高效的电池并不总是最稳定的。以此类推。

"The electrodes have to satisfy very different properties at the same time. They always conflict with each other," said Ritesh Kumar, an Eric and Wendy Schimdt AI in Science Postdoctoral Fellow working in the Amanchukwu Lab at the University of Chicago Pritzker School of Molecular Engineering (UChicago PME).
"电极必须同时满足截然不同的特性，而这些特性往往相互冲突，"芝加哥大学普利兹克分子工程学院（UChicago PME）Amanchukwu实验室的埃里克与温迪·施密特AI科学博士后研究员Ritesh Kumar表示。

Kumar is the first author of a new paper published in Chemistry of Materials that is putting artificial intelligence and machine learning on the job. The paper outlines a new framework for finding molecules that maximize three components that make an ideal battery electrolyte -- ionic conductivity, oxidative stability and Coulombic efficiency.
库马尔是一篇新论文的第一作者，该论文发表在《Chemistry of Materials》上，致力于让人工智能和机器学习发挥作用。论文概述了一个新框架，用于寻找能最大化理想电池电解质三个组成部分的分子——离子电导率、氧化稳定性和库仑效率。

Pulling from a dataset compiled from 250 research papers going back to the earliest days of lithium-ion battery research, the group used AI to tally what they call the "eScore" for different molecules. The eScore balances those three criteria, identifying molecules that check all three boxes.
该团队从250篇研究论文汇编的数据集中提取数据，这些论文可追溯至锂离子电池研究的早期阶段。他们利用人工智能计算了不同分子的"e Score"。e Score平衡了这三个标准，从而识别出符合所有三项条件的分子。

"The champion molecule in one property is not the champion molecule in another," said Kumar's principal investigator, UChicago PME Neubauer Family Assistant Professor of Molecular Engineering Chibueze Amanchukwu.
库马尔的主要研究者、芝加哥大学分子工程学院Neubauer Family助理教授Chibueze Amanchukwu表示："在某一特性中表现最佳的分子，在另一特性中未必最优。"

They've already tested their process, using AI to identify one molecule that performs as well as the best electrolytes on the market, a major advance in a field that often relies on trial-and-error.
他们已利用人工智能测试了该流程，成功识别出一种性能与市面最佳电解质相当的分⼦，这在依赖试错的领域是⼀项重大突破。

"Electrolyte optimization is a slow and challenging process where researchers frequently resort to trial-and-error to balance competing properties in multi-component mixtures," said Northwestern University Assistant Professor of Chemical and Biological Engineering Jeffrey Lopez, who was not involved in the research. "These types of data-driven research frameworks are critical to help accelerate the development of new battery materials and to leverage advancements in AI-enabled science and laboratory automation."
西北大学化学与生物工程助理教授杰弗里·洛佩兹表示："电解质优化是一个缓慢且具有挑战性的过程，研究人员经常需要反复试验来平衡多组分混合物中的竞争特性。"未参与该研究的洛佩兹指出："这类数据驱动的研究框架对于加速新型电池材料开发、利用AI赋能科学和实验室自动化进步至关重要。"

The music of batteries
电池的音乐

Artificial intelligence spots promising candidates for scientists to test in the lab so they waste less time, energy and resources on dead ends and false starts. UChicago PME researchers are already using AI to help develop cancer treatments, immunotherapies, water treatment methods, quantum materials and other new technologies.
人工智能为科学家筛选出实验室测试的有望候选方案，从而减少他们在死胡同和错误起点上浪费的时间、精力和资源。UChicago PME的研究人员已在利用AI协助开发癌症疗法、免疫疗法、水处理方法、量子材料及其他新技术。

Given that the theoretical number of molecules that could make battery electrolytes is 10 to the 60th power, or a one with 60 zeroes after it, technology that can flag likely winners from billions of non-starters gives researchers a huge advantage.
考虑到理论上可用于制造电池电解液的分子数量高达10的60次方（即1后面跟着60个零），能从数十亿种无效候选物中标记出潜力分子的技术为研究者提供了巨大优势。

"It would have been impossible for us to go through hundreds of millions of compounds to say, 'Oh, I think we should study this one,'" Amanchukwu said.
"要从数亿种化合物中筛选出‘哦，我觉得该研究这个’根本不可能，"阿曼楚库说。

Amanchukwu compared using AI in research to listening to music online.
Amanchukwu将研究中使用AI比作在线听音乐。

Imagine an AI trained on a particular person's musical taste, the combination of qualities that go into their own personal "eScore" for good songs. The new electrolyte research created the equivalent of an AI that can go through an existing playlist and, song by song, predict whether the person will like it. The next step will be an AI that can create a playlist of songs it thinks the person will like, a subtle but important conceptual tweak.
想象一个AI，它经过训练能理解某个人的音乐品味，即构成其个人“e Score”好歌标准的各种特质。这项新电解质研究相当于创造了一个AI，它能遍历现有播放列表，逐首预测这个人是否会喜欢这些歌曲。下一步将是开发一个能创建播放列表的AI，它会列出认为这个人会喜欢的歌曲，这是一个微妙但重要的概念调整。

The final step -- and the goal of the Amanchukwu Lab's AI research -- will be an AI that can write the music, or in this case design a new molecule, that meets all the parameters given.
最后一步——也是Amanchukwu实验室人工智能研究的目标——将是一个能够谱写音乐的人工智能，或者在这种情况下，设计出一种符合所有给定参数的新分子。

Amanchukwu last year received a Google Research Scholar Award to help the lab get closer to that final step: truly generative electrolyte AI.
Amanchukwu去年获得了Google Research Scholar Award，以帮助实验室更接近最终目标：真正无性向的电解质AI。

A quirk of graphic design
平面设计的怪癖

The team started curating the training data for the AI manually starting in 2020.
该团队从2020年开始手动为AI整理训练数据。

"The current dataset has thousands of potential electrolytes which we extracted from literature that spanned over 50 years of research," Kumar said.
"目前的数据集包含数千种潜在电解质，这些是我们从跨越50多年研究的文献中提取出来的，"Kumar说。

One of the reasons they have to enter the data manually comes not from chemistry, but from graphic design.
他们不得不手动输入数据的原因之一并非源于化学，而是源于平面设计。

When researchers write papers and journals lay them out in magazine format, the numbers the team turns into eScores are typically found in images. These are the jpeg or .png illustrations, charts, diagrams and other graphics that run within the text, but are not part of the text itself.
当研究人员撰写论文、期刊以杂志版面呈现时，团队转化为电子评分的数字通常出现在图像中。这些以JPEG或PNG格式呈现的插图、图表、示意图及其他图形虽内嵌于文本，却不属于文本本身内容。

Most large language models training with research papers just read the text, meaning the UChicago PME team will be manually entering training data for some time to come.
大多数大型语言模型在研究论文训练时仅读取文本，这意味着芝加哥大学PME团队在未来一段时间内仍需手动输入训练数据。

"Even the models today really struggle with extracting data from images," Amanchukwu said.
"即便是现在的模型，从图像中提取数据仍然非常困难，"阿曼楚库说。

Although the training data is massive, it's only the first step.
尽管训练数据庞大，但这只是第一步。

"I don't want to find a molecule that was already in my training data," Amanchukwu said. "I want to look for molecules in very different chemical spaces. So we tested how well these models predict when they see a molecule that they've never seen before."
“我不想找一个已经存在于训练数据中的分子，”阿曼丘克武说，“我想寻找化学空间迥异的分子。所以我们测试了这些模型在遇到从未见过的分子时预测效果如何。”

The team found that when a molecule was chemically similar to one from the training data, the AI predicted how good of an electrolyte it would make with high accuracy. It struggled to flag unfamiliar materials, marking the team's next challenge in the quest to use AI to design next-generation batteries.
团队发现，当某个分子与训练数据中的分子化学性质相似时，AI能以高准确度预测其作为电解质的性能。但该系统难以识别陌生材料，这标志着团队在利用AI设计下一代电池的任务中面临的下一个挑战。