Back to the roots at 50: Introducing Spectral-G4 (AI meets VLF)

[Taktyk]

Hi everyone,

Some of you might remember me from years ago, others probably not - it’s been quite a while since I was active in the detector building community.

I’m turning 50 in just a few days, and as I hit this milestone, I found myself reflecting on the projects that gave me the most joy over the years. I’ve spent my professional career as an AI Engineer in the IT industry, dealing with abstract algorithms and data. But lately, I’ve been missing the "real" engineering - the smell of solder, the challenge of a low-noise analog front-end, and the magic of physics.

I always remembered the Geotech1 community warmly, as a place of true innovation. So, I decided to return to my hobby, but this time, I’m bringing my professional "AI toolbox" with me.

I’ve been working on a project I call Spectral-G4. It’s a VLF platform that represents my vision of how modern tech can revitalize our hobby. Instead of the usual phase-ID, I’m implementing Neural Spectral Fingerprinting.

A few technical teasers of what I’ve been up to:

18-bit SAR High-Speed Acquisition: Using the AD7690 to get a crystal-clear signal.
20-bit Effective Resolution: Implementing hardware dithering and oversampling (a bit of an obsession of mine).
Distributed Intelligence: A Dual-MCU architecture with a "Smart Probe" (STM32G0) doing active vector nulling inside the coil.
The AI Layer: Using TinyML on an STM32G4 to classify targets based on harmonic profiles.

I’ve reached a stage where M1 (Analog/Power) is completed, and I’m now diving deep into the DSP and AI layers. I’ve set up a technical showcase and a landing page to document the journey.

I’m doing this primarily for the love of the craft, though I am considering a small "Founder Series" for those who want to join the development.

It feels great to be back. I’d love to hear your thoughts on the architecture.

Technical Showcase: https://github.com/takzen/spectral-g4-showcase

Project Website: https://spectral-g4.pl

Cheers,
Christopher

[Taktyk]

Hi everyone,

As promised, I have developed a detailed 34-point roadmap for the Spectral-G4 project. It covers the entire journey from the first solder joint in the analog stage to the final neural network inference.
I’m breaking the project down into five logical phases to track the progress more effectively. Here is the full plan:

Spectral-G4: Detailed Project Roadmap

Phase 1: Analog & Power Core (The Foundation)

1. TX H-Bridge Design: Designing a full H-bridge power stage for 7.8 kHz square-wave transmission.
2. MOSFET Driver Optimization: Implementing high-speed gate drivers for ultra-sharp signal edges (rich harmonics for AI).
3. Power Harvesting Circuit: Rectifying and filtering the TX signal to wirelessly power the in-coil electronics.
4. Ultra-Low Noise LDO Setup: Cascade power regulation (e.g., LT3042) with >80dB noise rejection for the analog rail.
5. Preamp Stage (AFE): Designing the initial ADA4841-2 based low-noise gain stage with precision bandpass filtering.
6. Differential ADC Driver: Implementing the ADA4941-1 to convert single-ended signals to fully differential for the SAR ADC.
7. High-Precision VREF: 4.096V voltage reference (REF3240) with <5ppm/°C drift.
8. Dithering Hardware: Implementing analog linear dithering via 47p/120p precision capacitors.

Phase 2: Distributed Intelligence (Smart Probe)

9. Probe PCB Layout: Miniaturized, vibration-resistant PCB for the coil (STM32G0 + Toroid).
10. I2C Protocol Bridge: Binary communication protocol between the Base (G4) and the Probe (G0) via a 6-wire cable.
11. 12-bit Active Vector Nulling: Implementing vector-based digital nulling using dual DAC channels inside the coil.
12. Toroid Summation: Calibrating the Al1800 toroidal transformer (10t:37t) for perfect zero-offset balancing.
13. Thermal Drift Monitoring: Firmware-level temperature sensing for dynamic phase-balance correction.
14. Coil ID System: Storing coil parameters (Model/Batch/Calibration) in the G0 Flash memory.

Phase 3: DSP & Data Acquisition (The Brain)

15. SPI+DMA High-Speed Link: Configuring STM32G4 SPI for low-latency AD7690 (18-bit) data acquisition.
16. HRTIM TX Generation: Utilizing the High-Resolution Timer for nanosecond-precision Dead-Time control.
17. CORDIC I/Q Demodulation: Hardware-accelerated phase and amplitude extraction using the CORDIC unit.
18. 20-bit Oversampling Engine: Software decimation and dither filtering to push the 18-bit SAR to 20-bit effective resolution.
19. Harmonic Extraction: Simultaneous demodulation of the 1st, 3rd, and 5th harmonics (7.8kHz, 23.4kHz, 39kHz).
20. EMI Noise Cancellation: Automatic frequency scanning to select the cleanest operating point.
21. Ground Balance Algorithm: Vector-based subtraction with automatic ground tracking logic.

Phase 4: AI & Neural Classification (TinyML)

22. Data Logging Tool: Building a raw spectral vector capture tool (Dataset Collector) via USB-VCP.
23. Neural Signature Mapping: Collecting a comprehensive dataset for non-ferrous, ferrous, and "trash" targets.
24. Model Architecture Design: Designing a lightweight MLP neural network optimized for the Cortex-M4 core.
25. X-CUBE-AI Integration: Porting the trained model (TensorFlow Lite) to optimized C code for the STM32G4.
26. Real-time Inference: Implementing on-device classification in the main detection loop (<10ms latency).
27. Static ID Logic: AI-assisted target identification for the Non-Motion (Static) mode.

Phase 5: Software, UI & Mechanics

28. Antigravity Framework Integration: Adapting low-level drivers to the Antigravity C++ framework.
29. GUI IPS Rendering: Designing the graphical interface for the IPS color display (spectral visualization).
30. Polyphonic Audio Engine: Low-latency VCO audio controlled by real-time AI classification parameters.
31. Mechanical Design: 3D-designed housing and ergonomic carbon fiber shaft.
32. Field Benchmarking: Comparative testing against commercial flagship detectors in high-mineralization soils.
33. B2B GIS Integration: Implementing data logging (VDI + Coordinates) for archaeological mapping.
34. Final Assembly & Calibration: Final calibration procedure for the "Series 0" Founder units.

I will be updating the GitHub repository as I progress through these stages. Feel free to ask technical questions about any specific step!

[EL NINO]

I salute you, Taktyk, it definitely looks like a very interesting project.. I'm definitely interested in it..and I see potential in it..

The technical specifications and parameters of the project themselves look really good...and it will be interesting to see what AI can do when learning on various test fields... and then in detection in the field..

[Taktyk]

Hi everyone,

To give you a better "bird's-eye view" of the project, I’ve put together a high-level block diagram of the Spectral-G4 architecture. I believe that in modern VLF design, the way we handle the signal before it even reaches the MCU is what separates a standard detector from a high-performance instrument.

Here’s a breakdown of the signal flow and the logic behind my choices:

1. The TX Engine (The Excitation)
I’m using a full H-Bridge driven by the STM32G4’s HRTIM (High-Resolution Timer). While many modern designs move towards Sine waves, I’m intentionally using a 7.8 kHz square wave with ultra-sharp edges. Why? To ensure a rich harmonic content (3rd and 5th harmonics) which are essential for the Neural Spectral Fingerprinting later in the chain.

2. The Smart Probe (Distributed Intelligence)
This is the heart of the "active" part of the project. Inside the coil, an STM32G0 acts as a co-processor.

• Active Vector Nulling: Instead of relying solely on mechanical geometry, I’m using dual DAC channels and a precision toroid transformer to inject a compensation signal. This allows for dynamic, firmware-controlled balancing of the RX/TX offset.
• Power: The probe is powered via a harvesting circuit from the TX lines, reducing the need for extra power wires in the cable.

3. The Analog Front-End (AFE)
Low noise is my obsession here.

• Preamp: An ADA4841-2 handles the initial gain.
• Differential Conversion: I’m using an ADA4941-1 to drive the ADC. Converting to a fully differential signal at this stage is crucial for EMI rejection and to fully utilize the SAR ADC’s range.
• VREF: A stable 4.096V reference (REF3240) ensures that my 18-bit measurements remain consistent across temperature shifts.

4. Data Acquisition & DSP
The signal is digitized by the AD7690 (18-bit SAR ADC) via a high-speed SPI/DMA link.

• Dithering & Oversampling: By adding controlled analog noise (dithering) and oversampling in the G4, I’m aiming for a 20-bit effective resolution.
• Hardware Acceleration: The CORDIC unit in the STM32G4 handles the I/Q demodulation in real-time, leaving the CPU cycles free for the AI layer.

5. The Neural Layer (The Decision Maker)
The demodulated vectors (1st, 3rd, and 5th harmonics) are fed into a TinyML model (TensorFlow Lite for Microcontrollers). Instead of a simple VDI number, the system outputs a classification probability based on the target’s spectral "fingerprint."

I’m particularly curious about your thoughts on the Active Vector Nulling via the toroid. Has anyone here experimented with active digital compensation in the coil assembly?

[Taktyk]

Following my introductory post, I’m diving straight into the first module of the Spectral-G4: the Transmitter (TX) Driver Stage.

In most VLF designs, the TX stage is often overlooked, but for a project relying on Neural Spectral Fingerprinting, the quality and "sharpness" of the excitation signal are everything. If we want to analyze the 3rd and 5th harmonics, we need a TX bridge that can switch fast and clean.

The Schematic Breakdown:

1. Full H-Bridge Architecture: I’ve opted for a full-bridge topology using two Si4599DY complementary N+P channel MOSFET pairs. This allows for a 24V peak-to-peak swing from a single 12V rail, maximizing the magnetic moment without complex power rails.
2. High-Speed Gate Driving: To minimize switching losses and, more importantly, to achieve ultra-sharp square-wave edges, I’m using the TC4427COA dual MOSFET driver. It provides up to 1.5A of peak current to quickly charge/discharge the gate capacitances.
3. Current Monitoring (The Feedback Loop): You’ll notice a 0.1R 1W shunt resistor (R12) at the bottom of the bridge. The signal TX_CUR_TO_STM32 (after a simple RC filter) is fed back to the STM32G4. This allows the system to:
   • Monitor real-time power consumption.
   • Detect over-current conditions (shorted coils).
   • Potentially implement a software-based resonance tuning check.
4. Bulk Decoupling: I’ve placed a 220uF low-ESR electrolytic capacitor (C47) in parallel with a 1uF ceramic (C17) right at the bridge entry. Switching several hundred milliamps at 7.8 kHz can create significant ripples; keeping the TX noise isolated from the future RX stage is priority one.

Why Square Wave?
While sine waves are "cleaner" for traditional phase-ID, they are energetically expensive to produce and lack the harmonic richness I need. By using a precisely timed square wave (via the G4’s HRTIM with nanosecond dead-time control), I’m effectively "illuminating" the target with multiple frequencies simultaneously.

Technical Question for the experts:
I am currently debating the optimal value for the shunt filter (R13/C18). I want to capture the current envelope without introducing too much phase lag for the protection logic. What are your thoughts on current sensing in high-speed TX bridges for VLF?

I've attached the KiCad schematic of the TX section. Looking forward to your thoughts!

Cheers,

[Carl-NC]

What's the purpose of R13/C18?

At White's I did some work on dynamic coil nulling but not with a transformer. I could make it work decently for a sine wave but wideband was far more difficult.

[profesor]

What's the purpose of R13/C18?

At White's I did some work on dynamic coil nulling but not with a transformer. I could make it work decently for a sine wave but wideband was far more difficult.

Hi Carl. Do you have a more detailed explanation about dynamic coil reset?

I'd like to try it. Thanks.

[Taktyk]

What's the purpose of R13/C18?

At White's I did some work on dynamic coil nulling but not with a transformer. I could make it work decently for a sine wave but wideband was far more difficult.

Hi Carl,

You’re right to point that out. To be specific: R13 and C18 are strictly for TX coil current monitoring in the main unit.

The transformer-based nulling you’re referring to is actually located inside the coil assembly (what I call the "Smart Probe"). It’s a separate subsystem where an STM32G0 MCU manages the active vector compensation using a small toroid for signal injection.

To provide a bit more technical context on how this works:

1. Toroid Summation: I am using a high-permeability toroid (secondary winding in series with the RX coil) as a summing transformer. The compensation signal is injected directly into the RX path at the source, before it enters the cable.
2. Local Vector Generation: The STM32G0 generates a precise compensation vector. By having the "brain" inside the coil, I can minimize the phase shifts and EMI pickup that usually occur when sending low-level analog compensation signals through a long cable.
3. 8-Core Connectivity: To ensure maximum stability, I am using an 8-core shielded cable. This allows for dedicated power lines (VCC/GND) for the probe electronics and a separate digital bus for communication between the G4 and G0, keeping the RX signal as clean as possible.
4. The "Saturation Prevention" Strategy: Addressing your point about wideband difficulty. I am not aiming for a perfect wideband zero. The primary goal is to null the 7.8 kHz fundamental. By removing the "carrier" physically at the source, I prevent the RX preamp (ADA4841) from saturating. This leaves the full dynamic range of the 18-bit SAR ADC available to capture the 3rd and 5th harmonics, which the AI then uses for classification.

In essence, I'm using the hardware to handle the "brute force" (the fundamental) and the AI to handle the "nuance" (the harmonic residuals).

I've attached the schematic of the Smart Probe below. I’d love to hear your thoughts on this approach, especially regarding the stability of the toroid injection over temperature.

[Taktyk]

I noticed a small error in the schematic I just shared regarding the transformer (T1). On the secondary side (injection), there are actually 10x fewer turns compared to the primary. This step-down ratio is crucial for better resolution of the compensation signal and lower output impedance.

I should also mention that this specific part of the architecture is inspired by commercial coils from the Polish manufacturer Rutus. In their designs, they implement phase and amplitude balancing using an AD5242 digital potentiometer.

For the Spectral-G4, I’ve decided to evolve this concept:

• Instead of a digital pot, I am using the MCU’s DAC channels.
• This allows for much finer granularity and faster dynamic adjustments during the nulling loop, which is essential when trying to maintain a stable null in varying soil conditions.

I’ve attached the corrected Schematic v1.1 and a photo of the original Rutus coil internal board for reference, showing their implementation of the digital pot balancing.

[Carl-NC]

Hi Carl. Do you have a more detailed explanation about dynamic coil reset?

I tried several approaches. I'll roughly describe them but without much detail because I may consider one for a future product.

One way is to use a small series resistor to get the TX current waveform, dynamically scale this signal, and then apply it either to the preamp or to the RX coil directly. A limitation of this method is that it works for a sinusoidal TX but not for a square wave TX.

Another way is to use differential bucking coils and dynamically drive them based on an error signal. This works with any TX waveform.

Another yet is to use a differential coil system like this

and dynamically adjust the balance drive to minimize Vout.

[Geotech11]

How is it progressing?

[pito]

You can use BTL amp, however from my experience BTL is giving double coil voltage but draws 5 times more current compared to a single amplifier.

some of my testings

[Taktyk]

How is it progressing?

Thanks for asking! The project is progressing intensely. I’ve just finalized the complete schematics for the Spectral-G4.
All major blocks are ready: the high-efficiency H-bridge for multi-harmonic TX, the ultra-low-noise Analog Front-End (AFE), and the integration of the 18-bit SAR ADC with the STM32G4.
Now I’m moving into the most critical phase: PCB Design. With 18-bit resolution and AI processing, signal integrity and a low noise floor are everything. I’m focusing on a multi-layer stackup to ensure perfect separation between the high-current TX paths and the sensitive RX AI-driven processing

[Taktyk]

Hi all,

Quick progress update - I'm officially starting on the software side of the Spectral-G4 (my multi-frequency VLF project with an AI "spectral fingerprint" discriminator). Hardware decisions are locked, so now it's time to write code.

--- The toolchain ---

I'm keeping it simple and modern. VS Code + the official STM32 VS Code extension is my whole IDE/debug setup for the STM32G474 (base) and STM32G071 (smart probe in the coil). The build system is CMake/Make driven by the extension (gcc-arm-none-eabi), flashing straight from the editor - no heavyweight IDE needed. For the number-crunching and signal design I use Python with numpy (and scipy for the solver), which lets me design and verify the DSP and the TX waveform on the PC *before* a single line goes onto the MCU. Nice separation: Python decides the math, the MCU just executes it.

--- Starting from the transmitter ---

Like any TX-first build, the first real question is: what exactly do we push into the coil? I don't want a plain 50% square wave - its harmonics are locked at 1, 1/3, 1/5, so the higher frequencies come out weak and you can't control them. I want THREE usable frequencies at the same time, all at comparable power, so the fingerprint has good dynamic range across the band.

The frequencies are odd harmonics of a base f0:

f0 = 7812.5 Hz (chosen as 250 kSPS / 32, so it grid-locks to the ADC)
-> 7.8125 kHz (1st), 23.4375 kHz (3rd), 39.0625 kHz (5th)

To make those three come out at EQUAL power from an H-bridge, I shape the switching pattern (SHE-PWM - selective harmonic shaping). I let numpy/scipy solve for the switching angles that force |b1| = |b3| = |b5| and kill the nearest contaminant (the 7th). The result I settled on (quarter-wave symmetric, so odd harmonics only):

switching angles: 51.47 deg / 64.62 deg / 85.33 deg
b1 = b3 = b5 = 0.5714 (equal - exactly what we want)
7th harmonic = 0 (nulled)

I cross-checked the analytic Fourier amplitudes against a brute-force FFT of the synthesized waveform - they agree to 4 decimals, so I trust it. This pattern becomes the HRTIM compare table on the G4; the timer just replays the edges (plus dead-time for the bridge).

One thing the math made obvious: hard-nulling MORE harmonics (7/9/11) is actually counterproductive - the energy just relocates into even higher harmonics, some right up near Nyquist. So killing only the 7th and letting the RX coil's bandwidth roll off the rest is the sweet spot. Lovely when the numbers tell you where to stop.

--- A small reflection ---

Honestly, what a time to be building this stuff. A free pro-grade editor, a full ARM toolchain, and Python+numpy to design and verify the signal chain on the bench before committing to silicon - all of it sitting on a laptop. The hobbyists who reverse-engineered these machines a couple of decades ago would have killed for tools like these. Great times to be a tinkerer.

--- Request to Carl ---

Carl - could you please move this thread over to the Projects section? I'd rather keep my build log there and not clutter up the main technical thread with progress posts. Thanks a lot, and thanks to everyone following along.

More soon - next up is getting the HRTIM pattern out on real pins and scoping the spectrum to confirm the three harmonics land at equal power.

Cheers

[Taktyk]

Before the G4: why I went back and rebuilt my old single-frequency machine first

A small update on the journey — and a confession about where the Spectral-G4 ideas are actually being battle-tested right now.

Before committing the G4 to copper, I made a decision that felt almost backwards: I went back to my old single-frequency VLF (a ~14 kHz IB machine on an 8-bit AVR that I built years ago) and rewrote its firmware from scratch. Not to sell it, not to modernize it for its own sake — but because I didn't want the G4 to inherit prototype-grade DSP. I wanted it to inherit proven DSP. The old machine became my lab rat: known coil, known analog front-end, no surprises — every behavior change is the algorithm's fault, nothing else's.

What got rebuilt, in rough order:

- Fixed-point everything. Q15/Q31 math and a CORDIC atan2 instead of floating point. The 8-bit AVR forces honesty here — if it runs there, the STM32G4 (with its hardware CORDIC) will eat it for breakfast.
- Full-phase target ID. The classic ratio tricks kept folding copper onto the ferrite side in my tests; a true atan2 over the demodulated X/Y vector, −90 (ferrite) to +90 (silver), turned out both simpler and more stable.
- Ground balance as an angle, not a coefficient. This one I owe directly to studying how the MXT was described by its designer: "resolution concentrated where high mineralization occurs." One knob maps through in the mineral-soil region, and the far end of the same scale reaches the salt domain (where the X channhe null lives atan absurdly large weight). No separate "salt circuit" — salt is just a balance poi
- A background ground analysis engine that tracks while you sweep
and halts when theet — again, an idea I shamelessly learned from the MXT's public description and
re-derived for my
- Profiles, EEPROM with versioned migrations (a firmware update
must never eat youSB telemetry stream that logs X/Y, ground-corrected residual, audio and balance at 50
Hz to CSV.

That last point caon of the wholeexercise, and I'll admit it cost me a broken afternoon: I tried to add overload detecerived from theory — datasheet full-scale, "reasonable" margins. Both attempts false-triggered cohain and deafenedthe machine. The fix wasn't a cleverer threshold; it was instrument first, ogger now existsprecisely so no constant in the G4 will ever be guessed again.
And that's the departure point. The rebuilt classic gives me one phase angle per ta a single frequency can produce. The Spectral-G4 exists because of what that number can't do: three hahases andamplitudes per target, and that vector is exactly the feature setthe neural classifion, sameground-tracking philosophy, same fixed-point modules ported almos1:1 — just instant once.
So: legacy code asgn philosophy asthe teacher, and the G4 as the thesis. The old machine goes into field testing nextthe G4 will have to beat.

— Christopher