Pioneering AI for system control and coordination
1 • Why use AI for system control
Hjarl solves constrained, non-convex optimization
- Exact solution to an imperfect representation: Linearizes equipment curves and constraints into a convex bowl.
- Strict modeling rules: Requires simplifying assumptions on dynamics and uncertainty.
- Hard to combine: Coupling adjacent systems rarely yields a convex problem.
- Imperfect solution to a near-exact representation: Captures non-linear dynamics, arbitrary distributions, and coupled systems.
- Complementary and improvable: Benchmarks alongside classical solvers and improves through iterative training.
- Testable optimum: Gradients (\nabla f) and shadow prices (\lambda^*) show what drives every action and how close a policy is to the optimum.
2 • Hjarl's offering
Local control over time, coordinated by signals in space
- Each policy optimizes control in its local domain over time and within hard limits.
- Pilots: Hydropower reservoir dispatch & direct air capture process control.
- Proof of concept: District heating boiler & storage control.
- Shadow price signals pass up and down the hierarchy so local decisions align with the global optimum.
- Showcase: Nodal shadow pricing in AC power grids.
3 • The Hjarl trinity
⅓ Engineering • ⅓ AI • ⅓ Optimization
Modular physics components run forward while exact gradients pass backward through time, replacing trial-and-error sampling.
Maps live system states and external drivers to control setpoints. Once trained, inference is immediate.
Updates control policies and constraint shadow prices (λ*) together so the system converges within hard physical limits.
Every decision is traceable through exact gradients and shadow prices.
4 • Not just for operators
For live advice, performance benchmarks, and system design
Read live state tags and return recommended setpoints and sensitivities as operating conditions shift.
Where the operator outperforms the policy, refine the simulator; where the policy outperforms, improve operating procedures.
Treat physical design parameters as trainable variables alongside the control policy to size and configure systems before committing capital.
5 • Path to running live · Removr / Honeywell DAC pilot
From simulator to sandboxed deployment
Assembled a transient thermodynamic simulator with adsorption hysteresis and kinetics transients.
Trained the policy for 400M simulated hours (10,000 × 40,000h batches) across weather, power price, and CO₂ price regimes.
Demonstrated meaningful economic uplift against the default operating policy, then deployed on a sandboxed server advising the Honeywell control system.
Example use cases
🌊 Hydropower dispatch
💨 Direct air capture
⚡ AC power grids
🏭 District heating
Optimal control use cases
🌊 Hydropower dispatch: Resource allocation under uncertainty

Modular waterway, generator, reservoir, and spillway simulator training constrained dispatch policies and terminal values (yielding reservoir water values ∂V/∂W) across >1-year horizons.
Avoids SDDP's simplifying assumptions on inflow distributions, price dynamics, and head dynamics - optimizing across multiple end markets with immediate inference.
- Co-optimization with batteries, pumping, and industrial loads.
- Load-based grid tariffs.
Optimal control use cases
💨 Carbon direct air capture: Thermodynamic process optimization

Default fan-speed rules optimize only for stability. Trained for 400M simulated hours (10,000 × 40,000h batches), showing meaningful economic uplift across weather and price swings.
- Inputs: Intake temperature (°C), relative humidity (%), power price (NOK/kWh), and CO₂ price (NOK/kg).
- Outputs & bounds: 0-100% fan speed, ending state value, and adaptive integration step within changeover and heat exchanger icing limits.
Runs on a sandboxed on-premise server with revokable certificates and no remote access, advising the Honeywell control system.
Optimal signals use case
⚡ Nodal power shadow pricing: Real-time markets for balancing
Solves AC power flow for nodal shadow prices and sensitivities (∂loss / ∂control). Built with general autograd rather than hardcoded derivatives, so it integrates directly with other optimizations.
Prices voltage, reactive power, active power, and line load bottlenecks in real time.
Click any bus or line on the map to pin an override, or toggle Settings to switch scenarios and sensitivity targets.
Proof of concept
🏭 District heating boiler & storage control


Coordinate a waste incinerator, oil/electric boilers, and storage buffers against volatile power prices and weather demand.
- Modeled the thermal grid as a differentiable simulator.
- Trained across 14,000 parallel simulations (~500,000 hours / 50 years).
Minimized fuel and peak-capacity tariff costs while meeting 100% of heat demand, using storage buffers optimally across cycles and seasons.
⚛ The core of Hjarl
What Hjarl delivers: A decision policy (π*)
An optimal operating line across shifting terrain
The core of Hjarl
What Hjarl delivers
A decision policyfrom an RL-trained agenttrained on a differentiable simulator
A function π*(s) mapping live states to control setpoints.Decision policy
A decision policy is an optimal operating line across shifting terrain
Maps live states to recommended control setpoints.
Measures the gap between historical operations and the policy.
Optimizes physical design parameters alongside control policies across scenarios.
🏃 Looking ahead
Pilot qualification & execution scope
Founder background
The Hjarl thesis at a glance
🏃 Let's build together
The path to running live
- Can be modeled in a differentiable simulator.
- Dynamics, distributions, or couplings break convex proxies.
- Continuous inputs and outputs.
- Synthetic scenario generation for external drivers.
- Forecasting models for uncertain inputs.
- System walkthrough and site visit.
- Historical series for external drivers.
- Domain contact for Q&A and benchmark review.
- Assemble the system simulator (
hjarl.systems). - Calibrate external drivers against historical data.
- Train the control policy (
hjarl.rl). - Deploy a prototype (
hjarl.serve+hjarl.dash). - Benchmark against historical operations.
Founder bio
👨🔧 Henrik Bakken
- 2000-05
- NTNU · M.Sc. Industrial Economics & Physics.Specialization in statistics and physics; master's thesis at the Norwegian Computing Center on copula theory (paper).
- 2005-08
- The Boston Consulting Group.Consultant.
- 2008-14
- SEB Enskilda Corporate Finance.Shipping & infra, director.
- 2014-19
- Nordic Capital.Private equity, director.
- 1988
- Gaming and tinkering.Got an Amiga in 1988.
- 2005
- Thesis on dependency structures.Realized that causality was vital to learning.
- 2007
- First differentiable simulator.Built to optimize gas flow in pipeline networks.
- 2015-19
- Nighttime AI researcher.Built CNN in 2015 (ResNet), LM in 2018 (BERT), custom CUDA kernels in 2019.
- 2020
- Founded Hjarl.A window to be part of solving learning.
In brief
The Hjarl thesis at a glance
- 1 • WHY
- Real physics isn't a smooth convex bowl.Classical solvers find the exact solution to an imperfect representation. Hjarl finds an imperfect, improvable solution to a near-exact representation.
- 2 • WHAT
- Local control over time, coordinated by signals in space.Policies optimize each asset over time within hard operating limits, while shadow prices coordinate actions across bottlenecks.
- 3 • HOW
- Modular physics simulation with differentiability at its core.The simulator runs physics forward and passes exact gradients backward, while an equilibrium solver updates policies and shadow prices together.
- 4 • FOR WHOM
- For live advice, performance benchmarks, and system design.Advise or automate live operations and benchmark historical decisions. Train physical design parameters alongside control policies ahead of investments.
- 5 • WHEN
- From simulator to sandboxed advisor.Assemble a system simulator, calibrate external drivers, and train a policy, then benchmark against historical operations before deploying in a sandbox.
- 6 • WHERE
- Tested on physical systems that can be differentiated.Hydropower dispatch (Skagerak Energi) and direct air capture (Removr / Honeywell) pilots, plus AC power grid and district heating showcases.
Appendix
RL training, simulation & differentiability (∇f)
Current state of AI control
Fusing RL with classical control theory
RL training
A self-learnt policy requires RL training
- Needs interactive rollouts; static logs are not enough.
- Sensitive to simulator accuracy and hard bounds.
- Unsafe to train directly on live industrial hardware.
- Learns multi-step strategies over long horizons.
- Handles continuous, coupled control and design variables.
- Purpose-built for physical control, unlike LLMs.
Simulation
Simulation makes RL feasible

- Logs only show past operator actions, not how the system responds to new policies.
- Past data goes stale when equipment, capacities, or rules change.
- Physical accuracy across operating regimes and bounds.
- High GPU throughput (10,000+ parallel rollouts).
- Modular code that evolves with the system.
Differentiability
Differentiability makes RL scalable
- Standard RL explores by random action noise.
- Required samples grow exponentially with control dimensions.
- Gradients show directly how each action affects the outcome.
- Replaces trial-and-error sampling with gradient descent.
- Cuts sample complexity by orders of magnitude.
Exact gradients scale policy learning to 100+ continuous control and design parameters.
Experimentally, our method shows superior asymptotic performance over MFRL baselines in complex locomotion tasks, achieving up to 64% higher reward even when baselines are given 10^6\times more training data. Further, AHAC’s efficient use of first-order gradients enables scaling to high-dimensional motor control tasks with 152 action dimensions.
Adaptive Horizon Actor-Critic (Nvidia Research)
Appendix: The current state of AI control
RL shows promise, but isn't fully researched

Complementary to classical solvers and improvable with iterative training, but without formal optimality guarantees.
Requires a simulator rather than static historical logs alone.
Nodal market making requires efficient bidders - one must first solve the local agents, and then solve the coordination.
Any sub-optimality can be inspected in the gradient - unlike black-box RL.
Appendix: Fusing AI with classical control
Fusing RL with classical control theory
- Differentiable physics replaces blind sampling with exact gradients.
- Example: First-order differentiable simulation (Nvidia AHAC) beats model-free RL by 64% with 10^6\times less data.
- Classical solvers (like SDDP) handle hard bounds by linking shadow prices (\lambda^*) to the value function.
- Hjarl brings shadow-price updates into policy learning over long horizons.
- Shadow prices and gradients make policies auditable (gradients vanish at the optimum).
- Extends classical principles to non-convex, coupled systems.
- Enforcing hard physical bounds.
- Optimizing over long horizons.
- Hydropower combines hard bounds, cascades, load-based tariffs, and >1y horizons.
- Serves as a concrete benchmark against SDDP.