发展基于“语义检测”的低参数量、多模态预训练电池通用人工智能模型

doi:10.19799/j.cnki.2095-4239.2024.0092

储能科学与技术 ›› 2024, Vol. 13 ›› Issue (4): 1216-1224.doi: 10.19799/j.cnki.2095-4239.2024.0092

• 电池智能制造、在线监测与原位分析专刊 • 上一篇下一篇

发展基于“语义检测”的低参数量、多模态预训练电池通用人工智能模型

吴思远¹^,²(), 李泓¹()

^1.中国科学院物理研究所北京清洁能源前沿研究中心
^2.中国科学院物理研究所凝聚态物质科学数据中心，北京 100190

收稿日期:2024-01-29 修回日期:2024-02-29 出版日期:2024-04-26 发布日期:2024-04-22
通讯作者: 李泓 E-mail:sywu@iphy.ac.cn;hli@iphy.ac.cn
作者简介:吴思远（1996—），男，博士，研究方向为固态电解质设计及第一性原理计算，E-mail：sywu@iphy.ac.cn；

Developing a general pretrained multimodal battery model with small parameters based on semantic detection

Siyuan WU¹^,²(), Hong LI¹()

^1.Institute of Physics, Chinese Academy of Sciences
^2.Condensed Matter Physics Data Center of Chinese Academy of Sciences, Beijing 100190, China

Received:2024-01-29 Revised:2024-02-29 Online:2024-04-26 Published:2024-04-22
Contact: Hong LI E-mail:sywu@iphy.ac.cn;hli@iphy.ac.cn

摘要/Abstract

摘要：

ChatGPT的出现意味着一种以“预训练+微调”为主的新科研范式诞生，以OpenAI为代表的企业正朝着训练通用人工智能（AGI）模型的路线前进，AGI意味着人工智能具备超越人类智力并解决通用性问题的能力，其是为了解决通用问题并具有强大的自学能力来促进人类社会发展。然而OpenAI等模型仍然是以文本为主结合图像等作为输入，对于电池体系来说，文本信息是少数的，更多的是温度、电压-电流曲线等的多模态数据，其所关注的结果包括电池荷电态、电池健康度、剩余寿命和是否出现电池性能跳水的拐点，甚至包括无数据情况下电池二次（梯度）利用的健康度评估。这意味着ChatGPT的路线虽然也可能解决电池体系的问题，但是以文本为主的样式或许有些“杀鸡用牛刀”，即使未来OpenAI的AGI可能解决当前电池存在的问题，但是在模型参数和输入方面与电池本质不符会使得模型参数量巨大而不适合电池离线端评估。对于电池体系的AGI，应该有自己独特的“文本语言”即理解电池运行过程中所发生的一切物理、化学过程以及其之间的关联，从而实现通用性并为后续全固态电池量产上车做铺垫。本文展望了在电池体系发展AGI过程中应该重新设计模型架构，特别在特征表示、数据结构设计、预训练方法、预训练过程设计和实际任务微调等需要重新设计。此外，相较于运行在服务器端的大模型，发展低参数量特别是离线的模型对于实时预测和基于我国国情及国际形势发展是十分必要的。本文主要讨论了发展基于“语义检测”的低参数量、多模态预训练电池通用人工智能模型所需要经历的几个阶段、可能面临的困难和评价指标，同时给出中国科学院物理研究所（以下简称物理所）在电池大模型在预训练、微调和测评三个方面“三步走”计划中的规划和可能线路。

关键词: 多模态, 通用人工智能, 电池状态, 语义检测, 预训练

Abstract:

The advent of ChatGPT signifies the birth of a new scientific research paradigm centered around "pretraining+fine-tuning". Companies such as OpenAI will lead the path toward artificial general intelligence (AGI) models. This indicates that artificial intelligence can surpass human intelligence and solve universal problems. AGI represents a model that is not designed for solving specific problems and even has the ability of self-learning. However, ChatGPT and other models still use texts combined with images as inputs. For a battery system, text information is less and most data as input is multimodal, such as temperature and voltage-current curve. The results related to a battery include the state of the battery including its charge and health, remaining useful life, whether there is a turning point in battery performance diving, and even the assessment of secondary (gradient) use of the battery without previous data. This means that ChatGPT can also help solve the battery system problem; however, its method can involve extra and complex solutions for minor problems even if AGI may solve the current battery problems in the future. Simultaneously, AGI can have huge parameters that are not suitable for offline operation of electric vehicles. We anticipate that AGI for a battery must have its own language and understand the physical and chemical processes during the operation of the battery. If AGI for batteries can understand why batteries become bad for example the lithium dendrites, they should predict all types of battery including all solid state battery. This review discusses how to redesign a battery model, including character representation, data distribution, pretrained methods and strategies, and fine-tuning for various tasks. In addition, minor parameters for the model should be concentrated for offline prediction and under international conditions. We will introduce the stages, problems, and evaluation indexes for developing a pretrained multimodal battery general model with minor parameters based on semantic detection (PBGM). We also present the three-step development strategy in PBGM by the Institute of Physics, Chinese Academy of Sciences (PBGM-IOPCAS).

Key words: multimodal, artificial general intelligence, pretrained, state of battery, semantic detect

中图分类号:

TP 391.77

吴思远, 李泓. 发展基于“语义检测”的低参数量、多模态预训练电池通用人工智能模型[J]. 储能科学与技术, 2024, 13(4): 1216-1224.

Siyuan WU, Hong LI. Developing a general pretrained multimodal battery model with small parameters based on semantic detection[J]. Energy Storage Science and Technology, 2024, 13(4): 1216-1224.

图/表 5

图1

图2

表1

图3

图4

参考文献 14

1	OpenAI: ChatGPT[EB/OL]. [2023-01-09]. https://openai.com/chatgpt.
2	DEVLIN J, CHANG M W, LEE K, et al. Bert: Pre-training of deep bidirectional transformers for language understanding[EB/OL]. [2023-12-06]. arXiv: .
3	RADFORD A, NARASIMHAN K, SALIMANS T, et al. Improving language understanding by generative pre-training[EB/OL]. [2023-01-09]. https://openai.com/research/language-unsupervised.
4	RADFORD A, WU J, CHILD R, et al. Language models are unsupervised multitask learners[ED/OL]. [2023-01-09]. https://cdn.openai.com/better-language-models/language_models_are_unsupervised_multitask_learners.pdf.
5	BROWN T B, MANN B, RYDER N, et al. Language models are few-shot learners[ED/OL]. [2023-07-06]. arXiv: .
6	BUBECK S, CHANDRASEKARAN V, ELDAN R, et al. Sparks of artificial general intelligence: early experiments with GPT-4[ED/OL]. [2023-07-06]. arXiv: .
7	MORRIS M R, SOHL-DICKSTEIN J, FIEDEL N, et al. Levels of AGI: Operationalizing progress on the path to AGI[ED/OL]. [2023-12-09]. arXiv: .
8	SULZER V, MOHTAT P, AITIO A, et al. The challenge and opportunity of battery lifetime prediction from field data[J]. Joule, 2021, 5(8): 1934-1955.
9	MOIN A, ZHENG Y, ANNA A, et al. The role of artificial intelligence in the mass adoption of electric vehicles[J]. Joule, 2021, 5(9): 2296-2322.
10	DUFEK E J, TANIM T R, CHEN B R, et al. Battery calendar aging and machine learning[J]. Joule, 2022, 6(7): 1363-1367.
11	MERCHANT A, BATZNER S, SCHOENHOLZ S S, et al. Scaling deep learning for materials discovery[J]. Nature, 2023, 624(7990): 80-85.
12	SINGH M, CAMBRONERO J, GULWANI S, et al. CodeFusion: A pre-trained diffusion model for code generation[ED/OL]. [2023-12-09]. arXiv: .
13	BURNS C, IZMAILOV P, KIRCHNER J H, et al. Weak-to-strong generalization: Eliciting strong capabilities with weak supervision[ED/OL]. [2023-12-16]. arXiv: .
14	PICARD D. Torch.manual_seed(3407) is all you need: On the influence of random seeds in deep learning architectures for computer vision[ED/OL]. [2023-12-09]. arXiv: .

等级	指标
1	传统机器学习模型，能预测单一体系的单一目标
2	通用模型构建，能够初步理解电池电化学反应，能够在跨体系预测、跨工况预测和跨尺度预测中一个达到设定的要求(见后文解释)
3	在2的基础上能够在跨体系预测、跨工况预测和跨尺度预测中两个达到设定的要求(见后文解释)
4	在2的基础上同时在跨体系预测、跨工况预测和跨尺度预测中达到设定的要求(见后文解释)
5	通用性大模型，在4的基础上不仅仅在二次电池能够达到设定的要求，同时可以扩展到一次电池、燃料电池等也有较强的泛化能力(见后文解释)

发展基于“语义检测”的低参数量、多模态预训练电池通用人工智能模型

Developing a general pretrained multimodal battery model with small parameters based on semantic detection

RichHTML

PDF

可视化

摘要/Abstract

引用本文

使用本文

图/表 5

参考文献 14

相关文章 0

编辑推荐

Metrics

本文评价