Choose one target with limited computational resources, such as an edge device used in a materials laboratory, a microscope control PC, a beamline auxiliary device, or a robot controller. For a classifier or detector that must run on that device, you can try exploration that includes 8-bit or lower precision conditions.
Examples suited to this include defect detection in microscope images, phase classification in XRD patterns, or anomaly signal detection during experiments, where latency and power consumption are important. It is good to begin by comparing only a few candidate architectures on a small dataset, and checking whether deployment-aware training is more stable than simple post-training quantization.
If the actual hardware is not yet fixed, you can first test in an environment that simulates the inference constraints of the target edge chip instead of CPU/GPU. The important point is to optimize for “operability on the actual device,” not for “performance-accuracy.”