While many models have been developed that can produce such probabilistic predictions, it is often the case that predictions are required for multiple related tasks, such that it would be desirable to have a probabilistic model that can make rapid predictions on new tasks without the need for task-specific training. Such is the case in the probabilistic meta-learning paradigm. While meta-learning has received an abundance of attention from the research community over the last decade (Finn et al., 2017; Gordon et al., 2019; Hospedales et al., 2022), the most notable class of probabilistic meta-model is, without doubt, the neural process family (NP; Garnelo et al., 2018a,b; Dubois et al., 2020). Recent advances in NPs have led them to reach astonishing heights in performance, representing the state-of-the-art in data-based approaches to weather and climate modeling (Bodnar et al., 2024; Allen et al., 2025; Ashman et al., 2024b), for example. Despite such impressive performance, industry practitioners seldom opt for deep learning models owing to their inherent lack of interpretability (Li et al., 2022), and instead prefer more traditional approaches such as kernel methods (Hofmann et al., 2008) that are easier to explain to non-technical stakeholders, even if they are incapable of meta-learning. Perhaps the most ubiquitous probabilistic model that practitioners turn to is the Gaussian process (GP; Rasmussen and Williams, 2005). With GPs, users can leverage their domain expertise to specify meaningful priors with which to bias predictions, any free parameters tend to have clear interpretations, and schemes such as automatic relevance T. Rochussen & V. Fortuin.

Environmental sensors are crucial for monitoring weather conditions and the impacts of climate change. However, it is challenging to place sensors in a way that maximises the informativeness of their measurements, particularly in remote regions like Antarctica. Probabilistic machine learning models can suggest informative sensor placements by finding sites that maximally reduce prediction uncertainty. Gaussian process (GP) models are widely used for this purpose, but they struggle with capturing complex non-stationary behaviour and scaling to large datasets. This paper proposes using a convolutional Gaussian neural process (ConvGNP) to address these issues. A ConvGNP uses neural networks to parameterise a joint Gaussian distribution at arbitrary target locations, enabling flexibility and scalability. Using simulated surface air temperature anomaly over Antarctica as training data, the ConvGNP learns spatial and seasonal non-stationarities, outperforming a non-stationary GP baseline. In a simulated sensor placement experiment, the ConvGNP better predicts the performance boost obtained from new observations than GP baselines, leading to more informative sensor placements. We contrast our approach with physics-based sensor placement methods and propose future steps towards an operational sensor placement recommendation system. Our work could help to realise environmental digital twins that actively direct measurement sampling to improve the digital representation of reality.