apex
Artificial Intelligence
for Policy Excellence in the Climate
Crisis
APEX empowers policy-makers to tackle the climate crisis by unlocking a much wider range of possible “futures” to explore. It builds on the deep domain knowledge already encoded in scientific simulators, using this foundation to develop Machine Learning solutions that are both robust and trustworthy. By accelerating computations, APEX makes it possible to evaluate many more scenarios in far less time, supporting better-informed decisions for a sustainable future.
APEX in short
The escalating impacts of climate change will reshape the physical and social environments humans rely on, making climate change adaptation and mitigation actions increasingly critical. Given the profound uncertainty, there exists a
multitude of potential environmental and societal outcomes. There is thus an unmet need for advanced tools that can dramatically extend decision-makers’ ability to comprehend
and respond to uncertainty.
The APEX Proof-of-Concept will seek to exploit the speed of Machine Learning solutions to allow the evaluation of a higher number of scenarios. This will be accomplished by making Machine Learning methods more robust out-of-distribution (e.g. in future scenarios characterized by scarce data) and thus more trustworthy. The methodological advancements will focus on embedding domain knowledge, physical laws and causal theory in the Machine Learning architecture.
This 5-year project aims at obtaining Machine Learning solutions able to mimic state-of-the-art simulators, beyond their training distribution without a substantial performance drop-off. This will allow their use in short computational times for scenarios of climate disruptions, with specific focus on the effects on the transport network. The APEX outcome will thus support real-world policy-making, providing a fast testing ground for key policy decisions in the face of climate disruption.
What’s the problem?
Climate-induced disruptions can assume many forms and result in multiple, undesirable futures. When interventions (e.g. policies) are implemented, the uncertainty increases and the number of scenarios to be evaluated becomes computationally intractable with traditional means. Simulators are usually used to assess different scenarios and compare different solutions, as they are able to frame multiple dimensions of the problem, the interactions between variables and the expected physical phenomena. Still, at large-scale, simulators require high computational times to be run, spanning over the time of weeks for each scenario. This constrain strongly reduces the amount of times a simulator can be used to evaluate a new policy and the amount of data that is ultimately available to the policy-makers to make an informed decision.
APEX envisions a data-informed framework that blends several simulation models using machine learning-based approximations, also known as metamodels. Traffic is chosen as main field of application, due to its connection to most of the other aspects of the urban system. If the developed metamodels can be improved in their performance while predicting unseen scenarios, they can reliably be used to inform policymaking greatly cutting the computational constraint. In harmonious interaction, these metamodels will create a consolidated, wide-ranging platform for testing key policy decisions impacting society, using a reinforcement learning framework as the backdrop.
Artificial Intelligence
To help uncover complex relations and while ensuring acceptable computational times
- Graph Neural Network
- Physic Informed Neural Network
- Causality
- Reinforcement Learning
Large scale simulations
To extract and embed domain knowledge, making machine learning solutions robust
- Traffic assignment (Aequilibrae, PTV VISUM)
- Climate (CICERO-SCM)
Support to policy-making
To explore how wider scenario exploration can inform policymaking and, in turn, lead to better social outcomes
- Reinforcement Learning
- Multi-Agent systems
Team
Current team
Meet the members of the team who are actively contributing to APEX:
Technical University of Denmark
Technical University of Denmark
Technical University of Denmark
Technical University of Denmark
Technical University of Denmark
Technical University of Denmark
Technical University of Denmark
Technical University of Denmark
Latest news
Sept., 2025
[+] Presentation at MT-ITS (9th IEEE Conference on Models and Technologies for Intelligent Transportation Systems). Oskar presented our work on meta-modeling the traffic assignment problem through GNN.
Abstract: The presented model is designed to mimic the structure used in conventional traffic simulators allowing it to better capture the underlying process rather than just the data. We benchmark it against other conventional deep learning techniques and evaluate the model’s robustness by testing its ability to predict traffic flows on input data outside the domain on which it was trained.
Scientific Presentation
Sept., 2025
[+] Poster presentation at FAST (Fast Machine Learning for Science). João presented our work on embedding physics into Neural Networks.
The work is an evaluation of the MixFunn architecture learning the Duffing oscillator. Different dynamics are fed to 2 MixFunn architectures and to a vanilla PINN and their performance out-of-distribution is benchmarked. The results show the potential of inductive biases to make neural networks faster (150x fewer FLOPs) and more robust out-of-distribution compared to vanilla PINNs.
Scientific Presentation
Sept., 2025
[+] Contributed talk at FAST (Fast Machine Learning for Science). Serio presented our work on building a taxonomy of inductive biases and how to use the resulting definition to transfer knowledge from simulators to Neural Networks.
Abstract: We hypothesise that simulators have already mathematically embedded domain knowledge, as a result of widely detailed physical phenomena and decades of development of dedicated algorithms. We analyze 3 main subjects: traditional inductive biases in ML and how they align with simulators; unconventional inductive biases inspired by the domain knowledge and generalizing power of simulators; algorithmic structures from the most common algorithms in large-scale simulators.
Scientific Presentation
July, 2025
[+] Contributed talk at UAI (41st Conference on Uncertainty in Artificial Intelligence). Francisco presented our work on Bayesian Hierarchical Invariant Prediction (BHIP): a scalable, prior-aware alternative to ICP for causal feature discovery.
We propose Bayesian Hierarchical Invariant Prediction (BHIP) reframing Invariant Causal Prediction (ICP) through the lens of Hierarchical Bayes. We leverage the hierarchical structure to explicitly test invariance of causal mechanisms under heterogeneous data, resulting in improved computational scalability for a larger number of predictors compared to ICP. Moreover, given its Bayesian nature BHIP enables the use of prior information. In this paper, we test two sparsity inducing priors: horseshoe and spike-and-slab, both of which allow us a more reliable identification of causal features. We test BHIP in synthetic and real-world data showing its potential as an alternative inference method to ICP.
Scientific presentation
Publications
1.
Learning traffic flows: Graph Neural Networks for Metamodelling Traffic Assignment
O.B. Lassen, Agriesti, S., Eldafrawi, M., Gammelli, D., Cantelmo, G., Gentile, G., & Pereira, F. C.
In 2025 9th International Conference on Models and Technologies for Intelligent Transportation Systems (MT-ITS). IEEE
Funding
This work is supported by the Novo
Nordisk Foundation grant NNF23OC0085356.
