Generated by eda-viz from the same divider block source.
| Layout view | Schematic view |
|---|---|
 |
 |
| Physical geometry with layer colors, ports, and routed metal. | Symbolic circuit view derived from the same Rust block. |
Generated by eda-waveform examples and gallery tools.
| Time-domain traces | Bode response |
|---|---|
 |
 |
| Clock + analog sample/hold traces with event timing context. | Magnitude and phase response for RC low-pass analysis. |
rlx-eda uses differentiable graphs plus autodiff to optimize both
model weights (surrogate training) and circuit parameters (inverse
design).
flowchart LR
A[Sample circuit design points] --> B[Ground truth target from circuit physics]
B --> C[Build rlx graph with MLP and loss nodes]
C --> D[Autodiff with grad_with_loss]
D --> E[Adam update]
E --> F{Converged?}
F -- No --> C
F -- Yes --> G[Trained surrogate parameters]
H[Inverse design target Vout at Vin] --> I[Build circuit loss graph]
I --> J[Autodiff wrt circuit params]
J --> K[Optimizer step on R1,R2]
K --> L{Converged?}
L -- No --> I
L -- Yes --> M[Found circuit parameters]
Losses used in the current pipeline:
- Surrogate training:
$L_{\text{surr}} = \frac{1}{B}\sum_{i=1}^{B}(\hat{y}_i - y_i)^2$ - Circuit inverse design:
$L_{\text{ckt}} = (V_{out} - V_{target})^2$
xychart-beta
title "Surrogate training loss (spike-surrogate, 1000 Adam steps)"
x-axis "step" [0, 100, 250, 500, 750, 999]
y-axis "loss" 0 --> 1.0
line [0.9858750, 0.2032971, 0.04010923, 0.03156146, 0.007113763, 0.003316347]
Measured run outputs (from this workspace):
| Optimization target | Initial parameters | Found parameters | Final metric |
|---|---|---|---|
Divider inverse design (spike-divider-block) |
R1=1000 Ω, R2=3000 Ω |
R1=2647.6 Ω, R2=1765.8 Ω |
Vout=0.400095 at Vin=1.0, loss 9.078e-9 in 151 iterations |
Surrogate training (spike-surrogate) |
Xavier init over MLP [3→16→1] |
81 learned weights/biases (W1,b1,W2,b2) |
loss from 9.858750e-1 to 3.316347e-3 over 1000 steps |
Single-circuit step-by-step trace (all optimization iterations):
cargo run -p spike-divider-block --bin ml_trace- see
../crates/spike-divider-block/docs/ml_optimization_trace_example.md