基于FTIR光谱检测分子中的功能组的二进制分类器的SHAP解释模型
Tomasz Urbańczyk1, Jakub Bożek2, Jarosław Koperski1
1Smoluchowski Institute of Physics, Faculty of Physics, Astronomy and Applied Computer Science, Jagiellonian University, Łojasiewicza 11, 30-348 Krakow, Poland.
像CNN-KAN这样的深度学习模型可以使用SHAP值来解释他们的决定来理解. 这项研究证实CNN-KAN对FTIR光谱的决定与已建立的化学组检测原则保持一致.
科学领域:
- 计算化学计算化学
- 频谱学是一种光谱学.
- 机器学习 机器学习
背景情况:
- 深度学习模型,特别是卷积神经网络 (CNN),提供了强大的功能,但经常充当"黑子",使其决策过程难以解释.
- 由于法律和道德方面的考虑,这种缺乏透明度可能会限制深度学习在受监管领域的采用,特别是在科学应用中.
- 解释模型决策的基础对于验证其可靠性和确保与科学理解保持一致至关重要.
研究的目的:
- 研究和阐明CNN-KAN (卷积神经网络 - 内核激活网络) 模型的决策过程.
- 为了确定福里埃转换红外光谱 (FTIR) 中哪些特定区域,CNN-KAN模型用于化学组识别.
- 评估模型的学习决策标准是否与科学文献中确定的化学原理保持一致.
主要方法:
- 一个CNN-KAN模型被训练为二进制分类器,使用FTIR光谱数据识别化学组.
- 为了分析和解释模型的预测,使用了夏普利增量扩展 (SHAP) 值.
- 使用SHAP值来确定影响模型分类决定 (正或负) 的特定光谱区域.
主要成果:
- 应用SHAP值成功追踪了CNN-KAN模型的决策路径.
- 特定的FTIR频谱区域被确定为积极和消极分类结果的关键驱动因素.
- 对模型决策负责的已识别的光谱区域与文献中已知对于检测特定功能组具有重要意义的区域相对应.
结论:
- 在识别FTIR光谱中的化学组时,CNN-KAN模型的决定是基于可解释的光谱特征.
- 该模型依赖于科学验证的光谱区域,提高了其在化学分析中的可信性和适用性.
- 这项研究证明了SHAP值在解密光谱学深度学习模型中的实用性,弥合了复杂算法和科学理解之间的差距.
更多相关视频
10:14Chromatographic Fingerprinting by Template Matching for Data Collected by Comprehensive Two-Dimensional Gas Chromatography
Published on: September 2, 2020
10:52Line Shape Analysis of Dynamic NMR Spectra for Characterizing Coordination Sphere Rearrangements at a Chiral Rhenium Polyhydride Complex
Published on: July 27, 2022
相关概念视频
IR Frequency Region: Fingerprint Region
Introduction to Functional Groups
Functional groups are group of atoms with specific chemical properties that occur within organic molecules and sometimes denoted as “R”. Functional groups are found along the carbon backbone of macromolecules can form chains or rings of carbon atoms. Functional groups can “functionalize” a compound by enabling it to adopt different physical and chemical properties.
Types of common functional groups
The table below summarizes some of the major functional groups in organic chemistry....
Molecular Shapes
Two regions of electron density in a diatomic...
IR Spectrum
Transmittance is defined as the ratio of the radiant power passing through a sample to that from the radiation's source. Multiplying the transmittance by 100 gives the percent transmittance (%T), which varies between 100% (no absorption) and 0%...
Molecular Models
Overview of Advanced Functional Groups
Functional groups are groups of atoms with specific chemical properties that occur within organic molecules and are sometimes denoted as “R”. Functional groups can “functionalize” a compound by enabling it to adopt different physical and chemical properties.
Types of Advanced Functional Groups
The table below summarizes some of the major functional groups in organic chemistry.
