Website NREL National Renewable Energy Laboratory (NREL)

Transforming energy!

The National Renewable Energy Laboratory (, is the nations only DOE laboratory dedicated to the research and advancement of renewable energy technologies.

The Complex Systems Simulation and Optimization Group in the NREL Computational Science Center has an opening for a full-time Postdoctoral Researcher – Computational Applied Mathematician for Reinforcement Learning Control and Optimization of Energy Systems. Artificial intelligence (AI) is rapidly becoming ubiquitous in renewable energy research. In particular, deep reinforcement learning (RL) for optimization and control of energy systems has the potential to transform our renewable energy future through advanced automation and control. We are looking for a dynamic researcher with a strong technical background to help pursue this goal. Stochastic optimization and sequential decision problems arise in a variety of areas of renewable energy research including power system planning and management, wind farm optimization, battery management, building energy management, transportation, even molecular design. At the same time, recent advances in RL suggest that complex decision making can be optimized via a combination of machine learning, modeling, and simulation, often with clear advantages over traditional optimization-based methods. Our initial efforts along these lines have opened the door to larger, sustained research programs in which an applied mathematician capable of exploring the deep underpinnings of designing and applying RL and multi-agent RL algorithms specifically tailored to NREL problems is necessary.

Direct responsibilities:

The successful candidate will collaborate to develop, adapt, improve, and scale cutting edge RL methods to real world projects in support of NREL and EERE mission and goals. The candidate will be tasked with core understanding of the optimal formulation of these algorithms. How deep RL approaches can be utilized in multi-agent control and optimization problems? How is existing domain knowledge to be utilized? How is complex parallel simulation code best incorporated into a scalable RL algorithm? What is the best way to combine traditional optimization and control algorithms and RL? We are looking for a researcher can to construct innovative RL algorithms from the ground up while appreciating domain-specific considerations. We are looking for a researcher who is able and excited to design novel RL approaches and bring the power of RL to pressing real-world problems (not just demonstrations and games).

Key areas of study include:

Model-based RL. There are a variety of efforts to use models in RL, but this is not a settled research area. We believe a combination of “top-down” incorporation of existing domain knowledge into “bottom-up” RL methods holds great promise, but this approach is not well developed.
Scalable RL. Supercomputers in principle allow for RL on many thousands of CPUs and hundreds of GPUs, yet existing algorithms and implementations do not yet take advantage of them. The successful candidate will participate in the development of scalable RL algorithms, especially for continuous and/or mixed spaces and that involve integration with complex parallel simulation code.
Hybrid methods. Classical optimization-based control such as Model Predictive Control (MPC) has many benefits, including interpretability and convergence guarantees, but (especially in real-time or for nonlinear systems) can become intractable. Methods that combine the best of both AI and classical stochastic optimization, e.g. that combine MPC with RL, are of great interest.
Multi-agent RL. Distributed optimization and control problems arise naturally in many clean energy applications. Methods that can address challenges in multi-agent RL (such as non-stationarity and credit assignment) and, more generally, enable distributed training and control, must be better understood in the context of clean energy applications.
Safe RL. For safety critical systems, there is well founded skepticism around using model-free RL as the primary controller due to challenges in handling complex system constraints and guaranteeing feasible control trajectories under uncertainty. Constraint aware, “safe” RL thus constitutes an important building block for bringing RL to bear on many real-world problems.
Evaluate and track the state of reinforcement learning research, especially for sequential decision problems in complex real-world spaces.
Author, present and assist in the preparation of technical papers, reports and conference proceedings on topics related to modern AI control methods and their application to energy systems.

To apply for this job please visit