AIC's 52nd Annual Meeting

This study focuses on the challenging task of identifying various species of mahogany, a prized wood sourced from the Caribbean in the 18th and early 19th centuries. Distinguishing between 'true mahogany' species and other tropical hardwoods, as well as North American woods mimicking mahogany, poses a significant challenge. Accurate wood identification is crucial for understanding the origins of raw materials, craftsmanship choices, and for effective conservation. Traditional methods involve microscopic examination by a wood anatomist, but obtaining suitable samples may not always be feasible or desirable. An alternative approach utilizes chemotaxonomy, leveraging variations in organic and inorganic chemical composition for wood differentiation.

In collaboration with Yale University Art Gallery, our ongoing project employs handheld laser-induced breakdown spectroscopy (LIBS) and pyrolysis gas chromatography-mass spectrometry (Py-GC/MS), complemented by machine learning (ML). Our goals are to distinguish mahogany from similar-looking woods and, ultimately, to differentiate between the three Swietenia mahogany species. Promising outcomes have emerged from the analysis of numerous samples, including those extracted from furniture. This presentation will highlight recent efforts to optimize data preprocessing steps, effectively deploy machine learning tools, and develop more robust classifiers. A collection of over 400 wood reference samples were studied with the two techniques, prior to examining approximately 200 areas on historic pieces of furniture employing the same approach.

Py-GC/MS is well-known among conservators for its efficacy in characterizing heritage materials. This method utilizes small samples of wood, either as tiny fragments or powdered material obtained with a hand drill. It takes about one hour to analyze a single sample in the laboratory. The resulting pyrograms show the presence of materials associated with cellulose, hemicellulose, and lignan, which are polymeric components common to all types of wood, as well as extractives, the non-structural, low molecular weight organic molecules that are the principal source of chemotaxonomic discrimination.

LIBS is a form of optical emission spectroscopy capable of simultaneously detecting all elements within a single laser pulse. Consequently, a broadband LIBS spectrum can be likened to a diagnostic fingerprint. With a commercially available handheld instrument, it is possible to analyze objects in situ in a matter of seconds. Notably, LIBS enables the detection of light elements, including both organic (e.g., C, H, O) and inorganic components (e.g., Li, Na, Mg, Al, Si, K, Ca, Ti, Fe, Zn, Sr).

Before applying machine learning tools, a preprocessing protocol was developed for the LIBS and Py-GC/MS data. This included baseline correction, alignment to ensure that the wavelength or retention time values were standardized across all data collections and were therefore directly comparable, and, in the case of LIBS, the removal of data with low signal-to-noise ratios (SNR) based on a spectral similarity analysis. Principal Component Analysis (PCA) and Partial Least Squares Discriminant Analysis (PLSDA) was then applied to build classifiers. Iterative refinement of ML algorithms and preprocessing steps resulted in models with a high level of classification success. Software was then developed to allow chemometric processing of LIBS data to be carried out in real-time in a gallery space.