Sicong Liu, Zhuoxian Zhang, Jie Zhang, Kecheng Du, Xiaohua Tong, Huan Xie, Yongjiu Feng, Yanmin Jin
{"title":"LIBSFormer:通过精确和可解释的氧化物定量,增强火星原位LIBS数据分析","authors":"Sicong Liu, Zhuoxian Zhang, Jie Zhang, Kecheng Du, Xiaohua Tong, Huan Xie, Yongjiu Feng, Yanmin Jin","doi":"10.1016/j.sab.2025.107204","DOIUrl":null,"url":null,"abstract":"<div><div>Accurate detection of the Martian surface composition is crucial for exploring signs of life on Mars. Since the Curiosity rover landed on Mars in 2012, Laser-Induced Breakdown Spectroscopy (LIBS) has been extensively utilized for in-situ analysis of Martian surface materials. However, multivariate methods employed for the quantitative analysis of LIBS data face challenges, including low feature extraction efficiency, insufficient interpretability, and poor generalizability. In this paper, we propose a novel Transformer-based approach for quantitative analysis of oxides in LIBS data, namely LIBSFormer. LIBSFormer employs the self-attention mechanism to automatically and holistically extract elemental features, associating information across the spectrum in a single step. Furthermore, we propose a wavelength-aware tokenization strategy for training, along with an element-driven approach for model selection, enhancing both model interpretability and reliability. The experimental results on the ChemCam LIBS RDR dataset demonstrate that LIBSFormer exhibits superior accuracy and stability in the quantitative analysis of Mars in-situ LIBS data, outperforming the state-of-the-art methods. The average Root Mean Square Error (RMSE) of LIBSFormer is 0.2199 wt%, with a Standard Deviation (SD) value of 0.0177 wt%. In comparison to the Convolutional Neural Networks (CNNs), LIBSFormer reduces the RMSE by 47.7 % and SD by 69.0 %, with similar results achieved on the ChemCam calibration dataset. With regard to interpretability, the attention weights of the feature extraction process have demonstrated that LIBSFormer is capable of effectively extracting key spectral lines with an accuracy of 90.0 %. These findings suggest that LIBSFormer has potential applications in high-accuracy quantitative analysis of Mars in-situ LIBS data.</div></div>","PeriodicalId":21890,"journal":{"name":"Spectrochimica Acta Part B: Atomic Spectroscopy","volume":"229 ","pages":"Article 107204"},"PeriodicalIF":3.8000,"publicationDate":"2025-04-03","publicationTypes":"Journal Article","fieldsOfStudy":null,"isOpenAccess":false,"openAccessPdf":"","citationCount":"0","resultStr":"{\"title\":\"LIBSFormer: Enhancing Mars in-situ LIBS data analysis with accurate and interpretable quantification of oxides\",\"authors\":\"Sicong Liu, Zhuoxian Zhang, Jie Zhang, Kecheng Du, Xiaohua Tong, Huan Xie, Yongjiu Feng, Yanmin Jin\",\"doi\":\"10.1016/j.sab.2025.107204\",\"DOIUrl\":null,\"url\":null,\"abstract\":\"<div><div>Accurate detection of the Martian surface composition is crucial for exploring signs of life on Mars. Since the Curiosity rover landed on Mars in 2012, Laser-Induced Breakdown Spectroscopy (LIBS) has been extensively utilized for in-situ analysis of Martian surface materials. However, multivariate methods employed for the quantitative analysis of LIBS data face challenges, including low feature extraction efficiency, insufficient interpretability, and poor generalizability. In this paper, we propose a novel Transformer-based approach for quantitative analysis of oxides in LIBS data, namely LIBSFormer. LIBSFormer employs the self-attention mechanism to automatically and holistically extract elemental features, associating information across the spectrum in a single step. Furthermore, we propose a wavelength-aware tokenization strategy for training, along with an element-driven approach for model selection, enhancing both model interpretability and reliability. The experimental results on the ChemCam LIBS RDR dataset demonstrate that LIBSFormer exhibits superior accuracy and stability in the quantitative analysis of Mars in-situ LIBS data, outperforming the state-of-the-art methods. The average Root Mean Square Error (RMSE) of LIBSFormer is 0.2199 wt%, with a Standard Deviation (SD) value of 0.0177 wt%. In comparison to the Convolutional Neural Networks (CNNs), LIBSFormer reduces the RMSE by 47.7 % and SD by 69.0 %, with similar results achieved on the ChemCam calibration dataset. With regard to interpretability, the attention weights of the feature extraction process have demonstrated that LIBSFormer is capable of effectively extracting key spectral lines with an accuracy of 90.0 %. These findings suggest that LIBSFormer has potential applications in high-accuracy quantitative analysis of Mars in-situ LIBS data.</div></div>\",\"PeriodicalId\":21890,\"journal\":{\"name\":\"Spectrochimica Acta Part B: Atomic Spectroscopy\",\"volume\":\"229 \",\"pages\":\"Article 107204\"},\"PeriodicalIF\":3.8000,\"publicationDate\":\"2025-04-03\",\"publicationTypes\":\"Journal Article\",\"fieldsOfStudy\":null,\"isOpenAccess\":false,\"openAccessPdf\":\"\",\"citationCount\":\"0\",\"resultStr\":null,\"platform\":\"Semanticscholar\",\"paperid\":null,\"PeriodicalName\":\"Spectrochimica Acta Part B: Atomic Spectroscopy\",\"FirstCategoryId\":\"92\",\"ListUrlMain\":\"https://www.sciencedirect.com/science/article/pii/S0584854725000898\",\"RegionNum\":2,\"RegionCategory\":\"化学\",\"ArticlePicture\":[],\"TitleCN\":null,\"AbstractTextCN\":null,\"PMCID\":null,\"EPubDate\":\"\",\"PubModel\":\"\",\"JCR\":\"Q1\",\"JCRName\":\"SPECTROSCOPY\",\"Score\":null,\"Total\":0}","platform":"Semanticscholar","paperid":null,"PeriodicalName":"Spectrochimica Acta Part B: Atomic Spectroscopy","FirstCategoryId":"92","ListUrlMain":"https://www.sciencedirect.com/science/article/pii/S0584854725000898","RegionNum":2,"RegionCategory":"化学","ArticlePicture":[],"TitleCN":null,"AbstractTextCN":null,"PMCID":null,"EPubDate":"","PubModel":"","JCR":"Q1","JCRName":"SPECTROSCOPY","Score":null,"Total":0}
LIBSFormer: Enhancing Mars in-situ LIBS data analysis with accurate and interpretable quantification of oxides
Accurate detection of the Martian surface composition is crucial for exploring signs of life on Mars. Since the Curiosity rover landed on Mars in 2012, Laser-Induced Breakdown Spectroscopy (LIBS) has been extensively utilized for in-situ analysis of Martian surface materials. However, multivariate methods employed for the quantitative analysis of LIBS data face challenges, including low feature extraction efficiency, insufficient interpretability, and poor generalizability. In this paper, we propose a novel Transformer-based approach for quantitative analysis of oxides in LIBS data, namely LIBSFormer. LIBSFormer employs the self-attention mechanism to automatically and holistically extract elemental features, associating information across the spectrum in a single step. Furthermore, we propose a wavelength-aware tokenization strategy for training, along with an element-driven approach for model selection, enhancing both model interpretability and reliability. The experimental results on the ChemCam LIBS RDR dataset demonstrate that LIBSFormer exhibits superior accuracy and stability in the quantitative analysis of Mars in-situ LIBS data, outperforming the state-of-the-art methods. The average Root Mean Square Error (RMSE) of LIBSFormer is 0.2199 wt%, with a Standard Deviation (SD) value of 0.0177 wt%. In comparison to the Convolutional Neural Networks (CNNs), LIBSFormer reduces the RMSE by 47.7 % and SD by 69.0 %, with similar results achieved on the ChemCam calibration dataset. With regard to interpretability, the attention weights of the feature extraction process have demonstrated that LIBSFormer is capable of effectively extracting key spectral lines with an accuracy of 90.0 %. These findings suggest that LIBSFormer has potential applications in high-accuracy quantitative analysis of Mars in-situ LIBS data.
期刊介绍:
Spectrochimica Acta Part B: Atomic Spectroscopy, is intended for the rapid publication of both original work and reviews in the following fields:
Atomic Emission (AES), Atomic Absorption (AAS) and Atomic Fluorescence (AFS) spectroscopy;
Mass Spectrometry (MS) for inorganic analysis covering Spark Source (SS-MS), Inductively Coupled Plasma (ICP-MS), Glow Discharge (GD-MS), and Secondary Ion Mass Spectrometry (SIMS).
Laser induced atomic spectroscopy for inorganic analysis, including non-linear optical laser spectroscopy, covering Laser Enhanced Ionization (LEI), Laser Induced Fluorescence (LIF), Resonance Ionization Spectroscopy (RIS) and Resonance Ionization Mass Spectrometry (RIMS); Laser Induced Breakdown Spectroscopy (LIBS); Cavity Ringdown Spectroscopy (CRDS), Laser Ablation Inductively Coupled Plasma Atomic Emission Spectroscopy (LA-ICP-AES) and Laser Ablation Inductively Coupled Plasma Mass Spectrometry (LA-ICP-MS).
X-ray spectrometry, X-ray Optics and Microanalysis, including X-ray fluorescence spectrometry (XRF) and related techniques, in particular Total-reflection X-ray Fluorescence Spectrometry (TXRF), and Synchrotron Radiation-excited Total reflection XRF (SR-TXRF).
Manuscripts dealing with (i) fundamentals, (ii) methodology development, (iii)instrumentation, and (iv) applications, can be submitted for publication.