An AI-controlled EV converter reached 96.7% on a lab bench—not in a vehicle
A 500 W prototype switched among five solar, battery and regenerative-braking modes. The neural controller itself was trained and tested on simulation data.
The short version
- What happened
- Researchers built a 500 W bidirectional converter prototype covering five power-flow modes between solar input, battery, load and regenerative braking.
- Who it affects
- An EV driver could benefit if a compact controller recovers more braking energy and manages solar or battery inputs efficiently. This study did not measure range, charging time, reliability or battery degradation in a vehicle.
- What remains unproven
- 500 W laboratory prototype
Keep in mind
500 W laboratory prototype
Five operating modes
Simulation-generated AI data
Single test split
Mechanical relay proof of concept
No road trial
No battery-lifecycle outcome
Source analysis · 1 linked source. Primary analysed source date: 29 October 2026.
Read the full report · 4 min The evidence, methods, implications and limitations
Five power-flow modes worked on a 500 W prototype
A laboratory converter combined photovoltaic input, a battery and regenerative-braking power flow in one bidirectional design, according to a peer-reviewed Frontiers in Artificial Intelligence paper published on 29 October. The 500 W prototype operated in five modes: battery-to-load, solar-to-load, combined solar and battery supply, solar charging the battery, and load-side regenerative charging. Peak measured efficiency was 96.7% at 400 W and remained above 92% from 100 W to 500 W.
The bench tests included loading from 100 W to 500 W, photovoltaic-source dropout and regenerative energy recovery. Reported transient response was below 2 milliseconds with overshoot no greater than 8%, and the MOSFET case stayed below 65°C at 500 W. These are useful proof-of-concept measurements for the power converter. They are not a claim about whole-vehicle range, charging time, battery life or energy saved on the road.[1]
The AI learned from simulation, not driving data
The controller used a modified deep neural network with one shared trunk and separate outputs for choosing the operating mode and setting switch duty ratios. Inputs included solar voltage and current, battery state of charge and voltage, load demand and regenerative-braking power. Its training data came from a MATLAB/Simulink model varied across all five modes, rather than from journeys, production batteries or long-duration hardware logs.
The network was compared with a baseline neural network, k-nearest neighbours, a bagged neural ensemble and an LSTM on the same 70:15:15 simulation-data split. The proposed model recorded the highest mode-classification accuracy and uniquely produced duty-ratio outputs. However, the accuracy figures were single-run test results, while only training times were repeated three times. The paper does not expose an independent physical-data evaluation of the controller's predictions.[1]
The distinction that matters
Evidence stops before a vehicle trial
Different parts of the system were tested at different levels of realism.
Controller
Train and test on simulation data
Single held-out split
Converter
Run a 500 W laboratory prototype
Five modes; 96.7% peak
Vehicle
Road and lifecycle testing
Not performed
Automotive validation is the missing step
The proof-of-concept used mechanical relays, with automotive solid-state selectors proposed for future high-speed duty. Real vehicles add vibration, temperature cycling, electromagnetic interference, component ageing, fast regenerative events, sensor faults and safety requirements that a 500 W bench cannot reproduce. The paper reports an emissions check up to 30 MHz, but not full automotive qualification, fault-injection coverage or lifecycle reliability.
The authors reported no funding and no commercial or financial conflicts. The assessment would change with a higher-power automotive-grade prototype, hardware-in-the-loop testing against unseen drive cycles, fault and thermal endurance tests, and a vehicle trial measuring energy recovery and battery stress against a conventional controller. For now, the result supports circuit feasibility and a simulation-trained control concept—not a deployment-ready EV efficiency gain.[1]