- Laser and beam stabilization
- AI applications such as drones and robotics in unfamiliar/unstable terrain
- Industrial plants with many sensors and controls, which need to use AI to improve throughput and efficiency
- Differential states training enables system identification on a free-drifting high input/output system without a knowledge of a mathematical model.
- In applications such as interferometric control on coherent beam combining, only a small fraction of the training dataset near the operating point is required, instead of mapping the entire parameter space. This is advantageous for training speed on large scale systems.
- The neural network can be continuously retrained using the locking data to capture slow transfer function change of the system, without introducing additional exploration actions or dithering.
- The neural network prediction is deterministic and fast. Inference speed of tens of nanoseconds can be achieved on an FPGA device in the case using the coherent combining demonstration.
Traditional controllers for unstable systems with many inputs and outputs usually cannot adapt to un-analyzed or partially analyzed systems. Generally, machine learning requires the system to be stable and reproducible during training, which is impractical in real applications due to system drift.
A typical iterative phase retrieval method derives laser phase information from the measured power, starting with a guess and converging following multiple iterations. However, this method is too slow for active feedback, and sensitive to measurement errors, thus presenting issues of scalability.
Researchers at Lawrence Berkeley National Laboratory have created a scalable, multi-functional machine-learning controller that can be used for large multi-input and output systems and unstable, changing environments.
The invention is a “double-state training scheme” where the machine learning controller learns differentially. The trained model is capable of building a differential map between the observation and the controller action and continuously predicting. The new method is robust against drift during training, and capable of continuous learning while operating. It automatically updates as the system changes, thus there is no need to re-train, and there is no need to stabilize the system otherwise while training.
The invention has broad application–anything that needs to learn in a complex environment that is varying while the device is being trained, with multiple inputs and outputs that need to be controlled. Examples include drones in windy conditions, laser beams, and industrial environments with multiple sensors.
Development Stage: TRL 7- Prototype
For More Information:
- Dan Wang
- Qiang Du
- Russell Wilcox
- Tong Zhou
- Christos Bakalis
- Derun Li
Status: Patent Pending
Opportunities: Available for licensing or collaborative research
See other technologies in this field: