By name: sensing-intelligence-trainable-metamaterial description: "Sensing Intelligence as a Trainable Metamaterial Property methodology. Optimize metamaterial body geometry via differentiable simulation to preprocess external stimuli, improving neural network sensing accuracy by up to 5x or reducing required sensors by 10x. Use when designing embodied sensing systems, neuromorphic perception pipelines, bio-inspired sensor optimization, or physical preprocessing for neural networks." license: Complete terms in LICENSE.txt metadata: arxiv_id: "2605.23967" published: "2026-05-26" tags: [embodied-sensing, metamaterial, differentiable-simulation, neuromorphic, physical-preprocessing, neural-network]
Sensing Intelligence as a Trainable Metamaterial Property
Core Concept
In biological systems, the body mechanically preprocesses stimuli before neural transduction — filtering, amplifying, and reshaping signals. This methodology treats body geometry as a learnable parameter by backpropagating sensing loss through differentiable physics simulation to optimize metamaterial structure for neural network interpretation.
Key Technical Insights
Trainable Body Intelligence: Rather than hand-designing physical preprocessing, the neural network trains its own body geometry via differentiable simulation — the sensing loss gradient flows through the physics simulator to update metamaterial design parameters.
Sensor Reduction: Optimized metamaterial bodies reduce required electronic sensors by nearly an order of magnitude (10x) while maintaining accuracy, by concentrating information into fewer measurement points.
Accuracy Enhancement: Across numerical and experimental scenarios, the optimized body improves sensing accuracy by up to 5x compared to naive physical designs.
Differentiable Simulation Pipeline: The core architecture chains: stimulus → metamaterial physics (differentiable) → sparse sensor array → neural network classifier, with end-to-end gradient flow enabling joint body-brain optimization.
Implementation Pattern
1. Define metamaterial geometry parameters θ_body
2. Set up differentiable physics simulator F(θ_body, stimulus)
3. Place sparse sensor array on metamaterial
4. Train neural network classifier on sensor readings
5. Backpropagate total loss L = L_task + λ·L_body through simulator
6. Update θ_body via gradient descent
Applications
- Neuromorphic perception: Design physical preprocessing for event cameras, tactile sensors
- Soft robotics: Optimize compliant body shapes for proprioceptive sensing
- Wearable BCI: Reduce electrode count through body-optimized signal concentration
- Acoustic sensing: Metamaterial lens design for source localization
- Structural health monitoring: Vibration-based damage detection with fewer sensors
Comparison to Traditional Approaches
| Approach | Body Design | Sensor Count | Accuracy |
|---|---|---|---|
| Manual + dense sensors | Hand-designed | High | Baseline |
| Manual + sparse sensors | Hand-designed | Low | Degraded |
| Trainable metamaterial | Learned | Low | Recovered |
Activation Keywords
sensing intelligence, trainable metamaterial, differentiable simulation, embodied perception, physical preprocessing, sensor optimization, neuromorphic sensing, body-brain co-optimization, mechanical preprocessing, sparse sensing