Reflectance spectroscopy remains a cornerstone of remote mineralogical analysis of airless bodies. However, traditional spectral unmixing techniques face limitations when dealing with nonlinear mixing, instrument-specific noise, or spectral coverage constraints. To overcome these, we have developed a suite of robust convolutional neural network (CNN) models trained to derive modal and chemical compositions of olivine-pyroxene-dominated surfaces, along with asteroid taxonomy classification. This abstract presents: (1) a concise overview of the model and its accessibility through a public web tool, (2) applications of the models to data from Itokawa, measured by the Hayabusa spacecraft, and (3) the inferred surface composition, space weathering state, and their correlation with topographic context. Forthcoming results will also include observations from the HyperScout-H imager, which is currently undergoing calibration. Model Overview and RobustnessOur neural network models were trained on laboratory reflectance spectra of olivine and pyroxene-rich samples, along with a combined dataset from DeMeo et al. (2009) and Binzel et al. (2019). The models predict:Relative volumetric mineral abundances (olivine, orthopyroxene, clinopyroxene), Compositional endmembers (e.g., fayalite and ferrosilite contents), Taxonomic class probabilities using a confidence-based scheme. To ensure robustness across instruments and spectral resolutions, the training dataset and models were systematically validated using spectra with varying resolution, wavelength coverage, transmission functions, and noise levels. The models achieved better than 10 percentage point precision in modal and chemical compositions (Korda et al., 2023a) and over 90% confidence in taxonomy classification (Korda et al., 2023b) when applied to high-quality input spectra (see Fig. 1). Comparable performance on realistic spectra with added noise and reduced spectral resolution (Korda and Kohout, 2024) confirms the models' reliability under observational constraints and supports their use for rapid compositional and taxonomical classifications. Figure 1: Model performance: composition predictions for various wavelength grids (left) and taxonomy classification as a confusion matrix (right). Web Tool for Public Use To support broader adoption of our spectral analysis models, we offer an online tool featuring a user-friendly web interface (Korda and Kohout, 2024), accessible at sirrah.pythonanywhere.com (see Fig. 2). Users can upload one or more reflectance spectra and apply optional preprocessing steps such as outlier removal, denoising, wavelength trimming, or the application of an instrument's spectral transmission function. Both raw and preprocessed spectra can be visualised and downloaded. Subsequently, users can apply the available machine-learning models for taxonomic classification and mineralogical composition—including models optimised for specific instruments such as HyperScout-H and ASPECT. All predictions are returned in downloadable CSV format. For data privacy, all uploaded input files are automatically deleted after processing. Figure 2: Web interface for uploading and analysing asteroid reflectance spectra. Users can apply preprocessing, visualise data, and run machine-learning models for taxonomy and composition. Correlations with Space Weathering and TopographyA key strength of the model is its ability to capture subtle spectral variations linked to space weathering processes. Unlike traditional methods that rely only on discrete spectral features, the model interprets the full spectral shape, including nonlinear mixing effects and weathering trends. In Korda et al. (2023b), we applied the models to spectra of Itokawa and observed a notable decrease in the olivine-to-pyroxene ratio in spectrally mature (weathered) regions. This aligns with olivine's lower resistance to space weathering: exposure to micrometeorite impacts and solar wind alters its crystalline structure, weakening its diagnostic spectral features. As a result, weathered surfaces appear relatively pyroxene-rich. Notably, our compositional predictions for Itokawa are in close agreement with laboratory analyses of returned samples, validating the model's physical accuracy and further strengthening confidence in its predictions. Surface composition and space weathering on Itokawa show a clear relationship with topography. Topographic highs, such as ridge crests and fresh crater walls, exhibit fresher material—likely due to recent exposure or regolith loss. In contrast, low-lying regions like valleys and depressions host older, spectrally mature regolith that has accumulated over time (see Fig. 3). These spatial patterns support the interpretation that downslope transport and regolith gardening are key processes shaping spectral evolution. The observed smooth transitions in spectral properties support a scenario of gradual weathering and ageing of surface materials. Taxonomically, we observe a smooth transition along the Q-S spectral sequence across Itokawa's surface. This sequence, widely recognised as a proxy for surface maturation, allows our model to infer the relative ages of different terrains. While absolute dating is beyond the scope of our models, it can be supplemented by the machine-learning model of Palamakumbure et al. (202X), which predicts absolute surface ages from reflectance spectra. Together, these models offer a comprehensive view of both compositional and temporal evolution. Figure 3: Predicted match score of the S-type asteroids on the surface of Itokawa. The numbers designate fresh and mature areas. Forthcoming Application to Mask and DeimosWe are currently preparing reflectance spectra from ESA's Hera spacecraft and its onboard HyperScout-H instrument, which observed both Mars and its moon Deimos. These datasets will soon be analysed using our neural network models, marking the first application of our method to Martian system targets. The analysis will investigate surface composition, space weathering effects, and geological diversity on both bodies, offering new insights into the origin and evolution of Deimos and the mineralogical context of the Martian surface. Preliminary results will be presented at the conference. ConclusionOur neural network-based framework offers a powerful and adaptable tool for interpreting reflectance spectra of airless bodies. It delivers detailed insights into surface mineralogy, chemical composition, and taxonomy, and is robust to variations in instrument characteristics. When applied to spatially resolved targets, the resulting taxonomy and composition maps can serve as proxies for geological processes, highlighting correlations with surface topography and weathering state. These capabilities support inferences about regolith dynamics, including downslope transport and resurfacing. The accompanying web interface ensures accessibility for the planetary science community and facilitates spectral interpretation for current and future missions. ReferencesBinzel et al. (2019): DOI 10.1016/j.icarus.2018.12.035DeMeo et al. (2009): DOI 10.1016/j.icarus.2009.02.005Korda et al. (2023a): DOI 10.1051/0004-6361/202243886Korda et al. (2023b): DOI 10.1051/0004-6361/202346290Korda and Kohout (2024): DOI 10.3847/PSJ/ad2685Palamakumbure et al. (202X): Under revision