Neural network model detection of tephra horizons in lake sediments using XRF elemental composition data

Tephra is a key tool for constructing stratigraphic and chronological frameworks that enable the precise correlation of palaeoclimate, palaeoenvironmental, and archaeological records. Cryptotephra, with its greater dispersal potential, can extend these frameworks to an intercontinental scale, which helps constrain rapid regional climatic and environmental transitions. Even very low-concentration tephra layers can play a significant role in refining these records, making it crucial to identify all tephra layers within a sediment record. Here, we aim to develop a novel method for detecting tephra within sediment records that is rapid, non-destructive, and capable of producing a comprehensive record.

Currently, cryptotephra layers are identified through laboratory methods, which can take anywhere from several months to years to process extensive lake sediment records (10s-100s of metres in length). To complement these methods, core scanning techniques (e.g. X-ray fluorescence (XRF), magnetic susceptibility, X-ray CT) are utilised to expedite the identification of tephra layers. Still, they can face limitations when identifying cryptotephra layers that share similar physical characteristics with the host sediment.

To build upon these identification methods, we have developed an AI neural network model. This model is trained on elemental compositions from a 1 mm resolution XRF scanning (AVAATECH) dataset, from the diatom and organic-rich varved sediment record of Lake Chala (Kenya/Tanzania). This model is designed to complement existing identification techniques by more efficiently predicting the presence of both visible and cryptotephra layers, surpassing the capabilities of standard statistical data reduction methods. It has undergone multiple training iterations and demonstrates the capability to predict all laboratory-identified tephra layers within a test subsection of the record. Additionally, the model has undergone sensitivity tuning to improve the accuracy of these predictions.

This model will enable more efficient screening of sediment cores and prioritisation of samples for laboratory analysis. By accelerating the detection process without compromising the completeness of the tephra record, the model supports the development of regional tephrostratigraphic frameworks, correlation of regional palaeorecords, and development of complete volcanic eruption records. Future work will focus on expanding the model’s applicability across diverse sedimentary records with different background-sediment compositions. We anticipate that this model will contribute to a more efficient and accessible approach to tephra detection in extended lake sediment records.