Building a mmWave Material Classification Radar

Building a mmWave Material Classification Radar

Overview

An engineer has developed a proof-of-concept (POC) millimeter-wave (mmWave) radar capable of classifying materials based on their electromagnetic signatures. The project aimed to solve the high cost and invasive nature of asbestos detection in European buildings by replacing physical lab sampling with non-destructive radar scanning.

Hardware Architecture

The prototype was built using a combination of off-the-shelf development boards to accelerate the prototyping phase and reduce initial hardware design risks.

  • Core Components: The system utilizes a Texas Instruments IWRL6432 BOOST for radar functionality and an ESP32 dev kit for control and connectivity.
  • Mechanical Design: The author adopted a "electronics-first" approach, designing 3D-printed mechanical casings around the existing development boards rather than designing a custom PCB from the start.

Digital Signal Processing (DSP) Chain

The radar operates on the Frequency Modulated Continuous Wave (FMCW) principle, where the transmitter emits a linear frequency sweep (chirp) rather than a single tone. The DSP chain converts these echoes into a material signature through the following stages:

  1. Chirp Generation and Characterization: The system characterizes the actual output sweep (start frequency, slope, and bandwidth) to ensure downstream accuracy.
  2. Mixing and Beat Signal: The received echo is mixed with the transmitted chirp, producing a low-frequency "beat" tone proportional to the distance of the reflector.
  3. Range FFT: A Fast Fourier Transform (FFT) converts the beat signal frequency into distance, providing reflected energy as a function of range (depth into the material).
  4. Capon Beamforming (AOA): Using a MIMO RX array, Capon (MVDR) beamforming resolves the Angle of Arrival (AOA) for each range bin, creating a sharp angular density spectrum.

The final output is a per-range, per-angle density "spectre"—an electromagnetic fingerprint that serves as the input tensor for the classification model.

Material Classification via AI

To classify the materials, the system employs a Convolutional Neural Network (CNN) that ingests the density spectre produced by the DSP chain. The model effectively learns the electromagnetic properties of materials, specifically the complex permittivity ($\epsilon'$ and $\epsilon''$).

Proof of Concept Results: The classifier was trained on a dataset containing 500 KB of spectral data per class across various environments and orientations. The system successfully classified combinations of materials, including:

  • Wood with aluminum, books, stone, or plastic.
  • Stone with aluminum, wood, books, plastic, or plexiglass.

RF Simulation and Reverse Engineering

To optimize antenna shapes and match the characteristics of the Texas Instruments hardware, the author used openEMS, an open-source FDTD (Finite-Difference Time-Domain) simulator.

Simulation Optimization Techniques:

  • Transfer Function Approximation: To avoid the high computational cost of full-wave simulations (which can take over an hour), the author calculated the transfer function from TX to RX using a Gaussian pulse instead of a Dirac pulse to maintain simulation stability.
  • Convolution: By convolving the chirp with the calculated transfer function, the simulation time was reduced from one hour to approximately two minutes.
  • Frequency Domain Analysis: The author used Bode diagrams to verify the simulated received signal against real measured outputs.

Commercial Challenges and Lessons Learned

Despite the technical success of the POC, the project ceased due to a lack of funding and market skepticism. Potential customers were unwilling to sign Letters of Intent (LOI) without a finished, regulation-cleared product in hand.

Key Engineering and Startup Lessons:

  • Market Validation: Validate willingness to pay via pre-orders and a landing page before building hardware.
  • **Lean Prototyping: Use development boards to quickly eliminate useless components and keep the design lean.
  • Design Order: Design mechanical casings around electronics, not vice versa, to avoid long PCB lead times.
  • Maintainability: Ensure products are OTA (Over-the-Air) updatable and maintain backward compatibility for code.

Community Insights and Counterpoints

Discussion surrounding the project highlighted several critical perspectives on the utility and the specific application of the technology:

"Asbestos doesn't work like that. If you don't touch it, you're fine... There's more harm from the fear of it."

Critics also questioned whether the POC actually proved the core value proposition—detecting asbestos shards within a material—rather than simply classifying different bulk materials.

"Unless I missed something, it seems the 'POC' device still made no effort to address this... it just demonstrated classifying some other common materials."

Other contributors suggested expanding the modality to detect discontinuities in material properties, which could potentially be applied to skin cancer detection or general-purpose industrial inspection.

Sources