Publications
Below is a full list of my publications. For an always up-to-date list, visit my Google Scholar profile.
Hjikakou, K., Cardenas-Cartagena, J., & Sabatelli, M. (2026). On the Generalisation of Koopman Representations for Chaotic System Control. Northern lights deep learning conference 2026.
This paper investigates the generalisability of Koopman-based representations for chaotic dynamical systems, focusing on their transferability across prediction and control tasks. Using the Lorenz system as a testbed, we propose a three-stage methodology: learning Koopman embeddings through autoencoding, pre-training a transformer on next-state prediction, and fine-tuning for safety-critical control. Our results show that Koopman embeddings outperform both standard and physics-informed PCA baselines, achieving accurate and data-efficient performance. Notably, fixing the pre-trained transformer weights during fine-tuning leads to no performance degradation, indicating that the learned representations capture reusable dynamical structure rather than task-specific patterns. These findings support the use of Koopman embeddings as a foundation for multi-task learning in physics-informed machine learning.
Assiotis, N., Hau, R., Oldenburg, V., Verbiest, R., Koellermeier, J., Sabatelli, M., & Cárdenas-Cartagena, J. (2025). Physics-Informed Graph Neural Networks for Air Pollution Forecasting in the Netherlands. ECAI workshop on machine learning meets differential equations: From theory to applications. PMLR, 47–70.
Accurate air pollution forecasting is critical for public health and environmental policy, particularly in densely populated regions like the Netherlands. This work introduces a physics-informed graph neural network (PI-GNN) framework for urban nitrogen dioxide (NO2) forecasting, which integrates domain-specific physical constraints into graph-based deep learning models. By combining spatial and temporal learning with physical knowledge, the proposed physics-informed graph convolutional network with gated recurrent units significantly outperforms purely data-driven recurrent and graph neural networks in terms of accuracy, generalizability, and environmental efficiency. Moreover, physics-informed models demonstrated progressively better relative performance over purely data-driven models in conditions with scarce data.
Kamsteeg, I., Cardenas-Cartagena, J., Beers, F. van, Tashu, T. M., & Valdenegro-Toro, M. (2025). Confidence Calibration in Large Language Model-Based Entity Matching. Proceedings of the 2nd workshop on uncertainty-aware NLP (UncertaiNLP 2025), 120–137.
This research aims to explore the intersection of Large Language Models and confidence calibration in Entity Matching. To this end, we perform an empirical study to compare baseline RoBERTa confidences for an Entity Matching task against confidences that are calibrated using Temperature Scaling, Monte Carlo Dropout and Ensembles. We use the Abt-Buy, DBLP-ACM, iTunes-Amazon and Company datasets. The findings indicate that the proposed modified RoBERTa model exhibits a slight overconfidence, with Expected Calibration Error scores ranging from 0.0043 to 0.0552 across datasets. We find that this overconfidence can be mitigated using Temperature Scaling, reducing Expected Calibration Error scores by up to 23.83%.
Lende, M. van der, Sabatelli, M., & Cardenas-Cartagena, J. (2025). Interpretable Function Approximation with Gaussian Processes in Value-Based Model-Free Reinforcement Learning. Northern lights deep learning conference 2025.
Estimating value functions in Reinforcement Learning (RL) for continuous spaces is challenging. While traditional function approximators, such as linear models, offer interpretability, they are limited in their complexity. In contrast, deep neural networks can model more complex functions but are less interpretable. Gaussian Process (GP) models bridge this gap by offering interpretable uncertainty estimates while modeling complex nonlinear functions. This work introduces a Bayesian nonparametric framework using GPs, including Sparse Variational (SVGP) and Deep GPs (DGP), for off-policy and on-policy learning. Results on popular classic control environments show that SVGPs/DGPs outperform linear models but converge slower than their neural network counterparts. Nevertheless, they do provide valuable insights when it comes to uncertainty estimation and interpretability for RL.
Müller, A., Cardenas-Cartagena, J., & Pollice, R. (2025). Uncovering Internal Prediction Mechanisms of Transformer-Based Chemical Foundation Models.
Supervised deep learning has become a standard approach to deliver competitive predictive tools that allow relating the structure of molecules and their physicochemical features to properties such as binding to protein targets, performance as electronic materials, and reactivity. However, efforts to understand how these models learn and how specific predictions can be explained are still limited in chemistry, which hinders trust in these predictions. To overcome both the current widespread approach of considering supervised deep learning models as black boxes and the limitations of explainable deep learning based on feature attribution, we introduce two general methods from the field of mechanistic interpretability to molecular property prediction. Specifically, we leverage ablation and adapt existing techniques into the regression lens to inspect model predictions of the Transformer-based ChemBERTa foundation model. Our results for 3 ChemBERTa-based models finetuned on distinct datasets allow us to propose the internal mechanisms operable within each of the corresponding model layers that lead to the final predictions. Our results are a stepping stone towards more trustworthy deep learning models in the molecular domain.
O’Cuilleanain, T., Cardenas-Cartagena, J., & Sabatelli, M. (2025). Adversarial Attacks Through Value-Guided Transition Modeling in Deep Reinforcement Learning. Northern lights deep learning conference abstracts 2026.
Efficient adversarial attacks on deep reinforcement learning agents rely on identifying critical states. Prior work uses learned transition models with environment-specific metrics to predict and lure the victim agent to such states. We propose a value-guided attack that integrates the victim policy’s value function as an environment-agnostic metric into both transition model training and state evaluation. From our preliminary results in the Pong environment from the Arcade Learning Environment, our method achieves comparable performance degradation to prior work while requiring roughly half as many attacks.
Todorov, A., Cardenas-Cartagena, J., Cunha, R. F., Zullich, M., & Sabatelli, M. (2025). Sparsity-Driven Plasticity in Multi-Task Reinforcement Learning. Transactions on Machine Learning Research.
Plasticity loss, a diminishing capacity to adapt as training progresses, is a critical challenge in deep reinforcement learning. We examine this issue in multi-task reinforcement learning (MTRL), where higher representational flexibility is crucial for managing diverse and potentially conflicting task demands. We systematically explore how sparsification methods, particularly Gradual Magnitude Pruning (GMP) and Sparse Evolutionary Training (SET), enhance plasticity and consequently improve performance in MTRL agents. We evaluate these approaches across distinct MTRL architectures (shared backbone, Mixture of Experts, Mixture of Orthogonal Experts) on standardized MTRL benchmarks, comparing against dense baselines, and a comprehensive range of alternative plasticity-inducing or regularization methods. Our results demonstrate that both GMP and SET effectively mitigate key indicators of plasticity degradation, such as neuron dormancy and representational collapse. These plasticity improvements often correlate with enhanced multi-task performance, with sparse agents frequently outperforming dense counterparts and achieving competitive results against explicit plasticity interventions. Our findings offer insights into the interplay between plasticity, network sparsity, and MTRL designs, highlighting dynamic sparsification as a robust but context-sensitive tool for developing more adaptable MTRL systems.
Cardenas-Cartagena, J., Falzari, M., Zullich, M., & Sabatelli, M. (2024). Upside-down Reinforcement Learning for More Interpretable Optimal Control. arXiv preprint arXiv:2411.11457.
Model-Free Reinforcement Learning (RL) algorithms either learn how to map states to expected rewards or search for policies that can maximize a certain performance function. Model-Based algorithms instead, aim to learn an approximation of the underlying model of the RL environment and then use it in combination with planning algorithms. Upside-Down Reinforcement Learning (UDRL) is a novel learning paradigm that aims to learn how to predict actions from states and desired commands. This task is formulated as a Supervised Learning problem and has successfully been tackled by Neural Networks (NNs). In this paper, we investigate whether function approximation algorithms other than NNs can also be used within a UDRL framework. Our experiments, performed over several popular optimal control benchmarks, show that tree-based methods like Random Forests and Extremely Randomized Trees can perform just as well as NNs with the significant benefit of resulting in policies that are inherently more interpretable than NNs, therefore paving the way for more transparent, safe, and robust RL.
Oldenburg, V., Cardenas-Cartagena, J., & Valdenegro-Toro, M. (2024). Forecasting Smog Clouds with Deep Learning: A Proof-of-Concept. ICML 2024 AI for science workshop.
In this proof-of-concept study, we conduct multivariate timeseries forecasting for the concentrations of nitrogen dioxide (NO2), ozone (O3), and (fine) particulate matter (PM10 & PM2.5) with meteorological covariates between two locations using various deep learning models, with a focus on long short-term memory (LSTM) and gated recurrent unit (GRU) architectures. In particular, we propose an integrated, hierarchical model architecture inspired by air pollution dynamics and atmospheric science that employs multi-task learning and is benchmarked by unidirectional and fully-connected models. Results demonstrate that, above all, the hierarchical GRU proves itself as a competitive and efficient method for forecasting the concentration of smog-related pollutants.
Cardenas-Cartagena, J., Beferull-Lozano, B., Elnourani, M., & Romero, D. (2022). Risk-Aware Particle Filtering for State Estimation in Recirculating Aquaculture Systems. 2022 asilomar conference on signals, systems, and computers, 829–833.
We design a sequential non-linear risk-aware estimator based on particle filtering to compute estimates and approximate the system state posterior distribution. For this purpose, we consider the risk given by the expected variance of the squared error between the system state and the estimate conditioned on the observations. We compare the proposed estimator with existing risk-neutral estimators in terms of error variance, mean squared error, and execution time per iteration. The comparison is carried out using a simulation of a Recirculating Aquaculture System, a case study for non-linear critical systems, where the performance in estimation for low probability events becomes an important aspect. Our simulation results demonstrate a competitive estimation performance while ensuring a lower risk.
Cardenas-Cartagena, J., Elnourani, M., & Beferull-Lozano, B. (2022). Forecasting Aquaponic Systems Behaviour with Recurrent Neural Networks Models. Proceedings of the northern lights deep learning workshop.
Aquaponic systems provide a reliable solution to grow vegetables while cultivating fish (or other aquatic organisms) in a controlled environment. The main advantage of these systems compared with traditional soil-based agriculture and aquaculture installations is the ability to produce fish and vegetables with low water consumption. Aquaponics requires a robust control system capable of optimizing fish and plant growth while ensuring a safe operation. To support the control system, this work explores the design process of Deep Learning models based on Recurrent Neural Networks to forecast one hour of pH values in small-scale industrial Aquaponics. This implementation guides us through the machine learning life-cycle with industrial time-series data, i.e. data acquisition, pre-processing, feature engineering, architecture selection, training, and model verification.
Hosamo Hosamo, H., Imran, A., Cardenas-Cartagena, J., Ragnar Svennevig, P., Svidt, K., & Kofoed Nielsen, H. (2022). A Review of the Digital Twin Technology in the AEC-FM Industry. Advances in Civil Engineering, 2022, 17.
The Architecture, Engineering, Construction, and Facility Management (AEC-FM) industry is increasingly affected by digital technologies that monitor sensor network data and control automation systems. Advances in digital technologies like Digital Twin offer a high-level representation of buildings and their assets by integrating the physical and digital world. This paper examines patterns, gaps, and trends in the AEC-FM sector and contributes to digitalization and automation solutions for building management. This work covers a broad range of research topics, from intelligent information management of complex models to building information management and the interaction of building systems, where researchers are increasingly interested in using the Digital Twin to manage their information and in developing new research lines focused on data interchange and the interoperability of building information modeling (BIM) and facility management (FM). After a complete bibliometric search of several databases and following selection criteria, 77 academic publications about the Digital Twin application in the AEC-FM industry were labeled and clustered accordingly. This study analyzed in detail the concept of key technologies, including “Digital Twin in Facility Lifecycle Management,” “Digital Twin-Information Integration Standards,” “Digital Twin-Based Occupants Centric Building Design,” “Digital Twin-Based Predictive Maintenance,” “Semantic Digital Twin for Facility Maintenance,” and “Digital Twin-Based Human Knowledge.” The findings show that information standardization is the first major hurdle that must be overcome before the actual use of Digital Twin can be realized in the AEC-FM industry. Based on that, this paper provides a conceptual framework of Digital Twin for building management as a starting point for future research.
Aguirre-Zapata, E., Cardenas-Cartagena, J., & Garcia-Tirado, J. (2017). Glycemic Monitoring in Critical Care Using Nonlinear State Estimators. IFAC-PapersOnLine, 50(1), 4430–4435.
As it is already known, patients under physiological stress, i.e., during a stay at an intensive care unit (ICU), often experience stress-induced hyperglycemia and insulin resistance even if they are not diagnosed with diabetes. In order to perform tight glycemic control (TGC) either by human intervention or by means of an automatic insulin infusion system, a proper monitoring system of the critical variables needs to be available. In this paper, the problem of estimating the state of the glucose-insulin system for a critically ill patient is investigated. Unlike most contributions, the interstitial glucose is assumed to be continuously measured by a continuous glucose monitor (CGM) instead of plasma glucose which is not usually available in most clinical facilities. Moreover, the measurement noise is characterized by a non-Gaussian noise as is suggested in recent studies. Three nonlinear filters are investigated, the Extended Kalman Filter (EKF), the Unscented Kalman Filter (UKF), and the Particle Filter (PF), under three different scenarios.
Cardenas-Cartagena, J., & Velásquez, V. H. J. (2017). Estimación de Posición En Robots móviles Usando Filtros de Partı́culas. Revista Politécnica, 13(25), 103–113.
This work presents an approach to solving the problem of controlling differential-motion mobile robots using odometry techniques, A*-based trajectory tracking algorithms, pure pursuit control, and particle filter state estimation to calculate the robot’s location. The paper includes a series of simulation results that verify the effectiveness of the proposed methodology.
Torres, M. S., Cardenas-Cartagena, J., Arenas, L. A., Quintero, P. A., & Torres, V. R. (2017). Cocreation Laboratory in Health: Materialization Tool for Innovation Process in Colombian Public Hospitals. 2017 congreso internacional de innovacion y tendencias en ingenieria (CONIITI), 1–6.
Although there is a considerable investment in research for overall health, only the 1% of Colombian hospitals and clinical institutions, known as IPS, can afford innovation processes, this phenomenon implies a short generation of intellectual assets in health-care system in Colombia is shown. To standardize innovation process, healthcare organizations must identify their capabilities to implement research, innovation and development activities; The General Hospital of Medellin (HGM) evaluated its own innovation capacities between 2010 and 2015, which resulted in the necessity to convert its strategic plan into innovative products and services. As result of this characterization, the HGM in alliance with the universities CES and EIA specialized in life science and engineering respectively, built an innovation strategy known as cocreation laboratory in health, which is understood as the place where engineering professionals, healthcare staff, and the community join forces to develop solutions prototypes to face in home hospital problems. The main objective of this strategy is to generate added value and increase the clinical and administrative hospital’s capacities, in order to gain quality in the patient’s attention.
Torres-Montoya, S., Cardenas-Cartagena, J., Torres-Villa, R., & Quintero-Posada, Á. (2017). Laboratorio de Cocreación En Salud. Encuentro internacional de educación en ingenierı́a ACOFI.
The implementation of a co-creation laboratory, following the MakerSpace model for healthcare areas at the Hospital General de Medellín “Luz Castro de Gutiérrez,” arises as a response to the need to generate innovation capabilities to solve organizational problems. In the first stage, challenges from clinical practice are addressed through the design of intra-hospital medical devices. This work presents the realization of the physical space, the strategy for collecting healthcare issues, and the weighting system used to prioritize them. During Phase I of the project, 74 problems were identified and categorized by innovation areas, from which the two most significant were prioritized. The laboratory aims to become the tool that materializes the innovation capabilities of public institutions within the healthcare sector.