Periodic Reporting for period 1 - UNCERTIR (Analytical Uncertainty of Quantitative Methods Based on IR-spectroscopy for Heritage Material Characterisation)
Reporting period: 2021-10-01 to 2023-09-30
Infrared (IR) spectroscopy combined with machine learning provided a step change in the quantitative determination of chemical and physical properties of many heritage materials of organic origin. However, for successful conservation, the characterisation needs to be accurate enough. Yet the uncertainty of IR spectroscopy-based methods has not had a due focus in the literature, leading to a reasonably limited acceptance of such methods by users. This Marie Skłodowska Curie Action, whose acronym is UNCERTIR, is of significant importance to heritage and users not only because of the variety of newly developed methods but also because, for the first time, the users will be provided with a comprehensive evaluation of uncertainty. Dating using (NIR) near-infrared spectroscopy and machine learning was of interest, so far significantly different accuracies have been observed for dating of paper, parchment or photographs. UNCERTIR specifically focused on the dating of real historical papers.
To provide a clear understanding of prediction uncertainties of NIR spectroscopy-based methods, especially for dating of paper, the research objectives were to quantitatively evaluate the possible sources of the uncertainty, which can be associated with historical material inhomogeneity, reference methods, spectra collection and chemometric approaches, and apply the results to a case study. A parallel goal of this Marie Skłodowska Curie Individual Fellowship was to foster the development of the researcher.
To provide a clear understanding of prediction uncertainties of NIR spectroscopy-based methods, especially for dating of paper, the research objectives were to quantitatively evaluate the possible sources of the uncertainty, which can be associated with historical material inhomogeneity, reference methods, spectra collection and chemometric approaches, and apply the results to a case study. A parallel goal of this Marie Skłodowska Curie Individual Fellowship was to foster the development of the researcher.
The work performed was ascribed to five work packages (WPs).
The intention of WP1 was to increase the Fellow’s research and soft skills through activities including day-to-day research work; training courses in line with the project aims and career prospects; meetings with colleagues, collaborators and advisory board members; visiting periods and secondments. WP2 sought to define the reference collection. Following a comprehensive literature review of IR spectroscopy-based methods for characterisation of heritage materials, dating of paper dated between 1851 and 2000 was chosen as of particular interest, mainly for two reasons (i) this is where data with a high degree of accuracy can be collected from the books themselves and/or archival catalogues, and (ii) significantly different accuracies have been observed so far, the cause of which was not known. The database which corresponded to the completion of WP2 fed into WP3 and WP4. While WP3 aimed to explore the contribution of material composition and reference method to the overall uncertainty, WP4 investigated the contributions of the methods of spectra collection and chemometric approaches. Within WP5, the overall method uncertainty for the dating of paper in the period 1851-2000 was evaluated in the case study carried out at the National and University Library of Slovenia (Ljubljana). To have a representative sample set, a total of 100 books were analysed according to a stratified random sampling, with the decade of publication as the criterion for stratification. During the period of relevance (1850−2000), the co-presence of rosin-sized paper, bleached paper, and quasi-neutral and alkaline paper with various cationic starching techniques can be found. Moreover, the same book can be made of paper from different batches. Therefore, NIR spectra were acquired on different pages of the books. In addition, since inhomogeneity can be due to degradation products, different parts of the page were analysed (e.g. inner and outer margins, the latter expected to be more degraded as exposed to pollutants and light). Several preprocessing techniques, including spectral preprocessing and variable selection, were tested, as well as several machine learning methods. The accuracy of the predicted data was used to compare different preprocessing strategies and machine learning methods, and to understand whether and how paper variability and natural degradation influence the dating models. It was shown that common spectral features of cellulose and protein structures are of importance for the predictions, rather than degradation that does not meaningfully influence the prediction accuracy. An unprecedented accuracy of as much as 2 years was achieved, better than any other non-destructive method applied to a real heritage collection. Results and further details, as well as spectra acquired, were published in “Coppola F., Frigau L., Markelj J., Malešič J., Conversano C., Strlič M. Near-Infrared Spectroscopy and Machine Learning for Accurate Dating of Historical Books. Journal of the American Chemical Society, 2023 145 (22), 12305-12314. DOI: 10.1021/jacs.3c02835”.
Throughout the project, to ensure that the research objectives were met, the project management included data, time, resource and risk management.
The results of UNCERTIR were disseminated and communicated with the intent to reach the academic and scientific audience, as well as the heritage institutions and the general public. The results were published in the top-tier Journal of the American Chemical Society “Coppola F., Frigau L., Markelj J., Malešič J., Conversano C., Strlič M. Near-Infrared Spectroscopy and Machine Learning for Accurate Dating of Historical Books. Journal of the American Chemical Society, 2023 145 (22), 12305-12314. DOI: 10.1021/jacs.3c02835”. Throughout the project, the Fellow was engaged in several communication and dissemination activities, including international conferences, lectures, media interviews and talks, not only about the aims and results of UNCERTIR, but also about the Fellow’s experience in writing and performing a project in such competitive fellowships, like the Marie Skłodowska Curie Action.
The intention of WP1 was to increase the Fellow’s research and soft skills through activities including day-to-day research work; training courses in line with the project aims and career prospects; meetings with colleagues, collaborators and advisory board members; visiting periods and secondments. WP2 sought to define the reference collection. Following a comprehensive literature review of IR spectroscopy-based methods for characterisation of heritage materials, dating of paper dated between 1851 and 2000 was chosen as of particular interest, mainly for two reasons (i) this is where data with a high degree of accuracy can be collected from the books themselves and/or archival catalogues, and (ii) significantly different accuracies have been observed so far, the cause of which was not known. The database which corresponded to the completion of WP2 fed into WP3 and WP4. While WP3 aimed to explore the contribution of material composition and reference method to the overall uncertainty, WP4 investigated the contributions of the methods of spectra collection and chemometric approaches. Within WP5, the overall method uncertainty for the dating of paper in the period 1851-2000 was evaluated in the case study carried out at the National and University Library of Slovenia (Ljubljana). To have a representative sample set, a total of 100 books were analysed according to a stratified random sampling, with the decade of publication as the criterion for stratification. During the period of relevance (1850−2000), the co-presence of rosin-sized paper, bleached paper, and quasi-neutral and alkaline paper with various cationic starching techniques can be found. Moreover, the same book can be made of paper from different batches. Therefore, NIR spectra were acquired on different pages of the books. In addition, since inhomogeneity can be due to degradation products, different parts of the page were analysed (e.g. inner and outer margins, the latter expected to be more degraded as exposed to pollutants and light). Several preprocessing techniques, including spectral preprocessing and variable selection, were tested, as well as several machine learning methods. The accuracy of the predicted data was used to compare different preprocessing strategies and machine learning methods, and to understand whether and how paper variability and natural degradation influence the dating models. It was shown that common spectral features of cellulose and protein structures are of importance for the predictions, rather than degradation that does not meaningfully influence the prediction accuracy. An unprecedented accuracy of as much as 2 years was achieved, better than any other non-destructive method applied to a real heritage collection. Results and further details, as well as spectra acquired, were published in “Coppola F., Frigau L., Markelj J., Malešič J., Conversano C., Strlič M. Near-Infrared Spectroscopy and Machine Learning for Accurate Dating of Historical Books. Journal of the American Chemical Society, 2023 145 (22), 12305-12314. DOI: 10.1021/jacs.3c02835”.
Throughout the project, to ensure that the research objectives were met, the project management included data, time, resource and risk management.
The results of UNCERTIR were disseminated and communicated with the intent to reach the academic and scientific audience, as well as the heritage institutions and the general public. The results were published in the top-tier Journal of the American Chemical Society “Coppola F., Frigau L., Markelj J., Malešič J., Conversano C., Strlič M. Near-Infrared Spectroscopy and Machine Learning for Accurate Dating of Historical Books. Journal of the American Chemical Society, 2023 145 (22), 12305-12314. DOI: 10.1021/jacs.3c02835”. Throughout the project, the Fellow was engaged in several communication and dissemination activities, including international conferences, lectures, media interviews and talks, not only about the aims and results of UNCERTIR, but also about the Fellow’s experience in writing and performing a project in such competitive fellowships, like the Marie Skłodowska Curie Action.
UNCERTIR has pushed the frontiers of IR spectroscopy combined with chemometrics forward in numerous ways. The potential of such methods for the characterisation of heritage objects, especially for paper dating, was explored with very promising results. The importance and role of a systematic evaluation of the uncertainty associated with such methods and of the underlying principles, so far neglected and without which broader acceptance of IR spectroscopy-based quantitative methods cannot be achieved, were emphasized. In particular, the approach and results elaborated will contribute significantly to the development of future systematic investigations for the determination of properties of heritage materials of organic origin, using IR spectroscopy and machine learning. Indeed, the methodology that was developed within UNCERTIR project is now standard in ongoing research to date more ancient paper. More generally, such investigations would increase the practical applicability of non-destructive IR spectroscopy-based quantitative methods to collections in heritage institutions, and enhance dozens of publications in the coming years.