This example documents nine iterations of analytical model refinement for a five-transistor OTA. Model predictions are compared with HSPICE measurements at each iteration, and the converged scalar model is converted to a vectorized evaluator.
Iterative equation refinement using HSPICE measurements
The driver normally writes to 5t_ota/. This snapshot is stored as 5t_ota_example_run/ so publishing it does not confuse it with a live workspace. The driver deletes an existing WORK_DIR before every run.
01
Extract circuit topology
The DUT netlist is reduced to the device and node information required to formulate topology-dependent equations and gm/Id constraints.
02
Generate analytical equations
Each iteration produces a structured set of performance equations. The equations are recorded and compiled into an executable Python model.
03
Perform LUT-based sizing
The scalar model and gm/Id lookup table generate device dimensions and the iteration's .1 prediction report.
04
Run HSPICE characterization
The selected device parameters are applied to the testbenches. Extracted measurements are stored in the corresponding HSPICE report.
05
Evaluate with operating points
HSPICE operating-point parameters are supplied to the analytical equations to produce the model values used in the comparison.
06
Calculate relative error
Model predictions are compared with HSPICE measurements. Equations exceeding the tolerance are revised in the next iteration; iteration 9 meets the convergence criterion.
Archived data structure
Files produced during model refinement
Each directory in the published snapshot has a local README. The table below shows its place in the dataflow.
netlist/
Topology context
The compact six-transistor DUT used by initial and feedback prompt generation.
Each cell below is the absolute relative error between the analytical model and HSPICE for that metric and iteration. Green cells pass the 5% accuracy limit; red cells fail it. DC gain remains at 1.46% because it passed in Iteration 1 and the refinement flow preserves equations that already pass while correcting only failed metrics.
Metric
Iteration 1
Iteration 2
Iteration 3
Iteration 4
Iteration 5
Iteration 6
Iteration 7
Iteration 8
Iteration 9
DC gain
1.46%
1.46%
1.46%
1.46%
1.46%
1.46%
1.46%
1.46%
1.46%
3 dB bandwidth
4.85%
4.85%
4.85%
4.85%
4.85%
4.85%
4.85%
4.85%
4.85%
UGB
1.53%
1.53%
1.53%
1.53%
1.53%
1.53%
1.53%
1.53%
1.53%
Phase margin
0.25%
0.01%
0.01%
0.01%
0.01%
0.01%
0.01%
0.01%
0.01%
ICMR min
6.74%
11.62%
6.02%
5.15%
2.86%
2.86%
2.86%
2.86%
2.86%
ICMR max
4.37%
4.37%
4.37%
4.37%
4.37%
4.37%
4.37%
4.37%
4.37%
Output swing min
141.34%
3.73%
3.73%
3.73%
3.73%
3.73%
3.73%
3.73%
3.73%
Output swing max
8.08%
8.08%
8.08%
8.08%
8.08%
8.08%
1.62%
1.62%
1.62%
Positive slew
37.59%
24.82%
17.71%
21.89%
18.91%
17.73%
44.70%
21.89%
0.72%
Negative slew
32.90%
101.30%
32.90%
4.12%
4.12%
4.12%
4.12%
4.12%
4.12%
PSRR+
15.21%
10.57%
8.40%
9.33%
5.88%
1.75%
1.75%
1.75%
1.75%
PSRR−
15.14%
3.71%
3.71%
3.71%
3.71%
3.71%
3.71%
3.71%
3.71%
CMRR
63.43%
16.45%
19.12%
5.62%
57.62%
10.40%
5.62%
3.12%
3.12%
01
Initial model
Model gain19.930 dBHSPICE gain20.226 dBGain error1.46%Worst error141.3%
Correct (5): DC gain, bandwidth, UGB, phase margin, ICMR max.
Needs work (8): ICMR min, both output swings, both slew rates, PSRR+, PSRR−, CMRR.
Largest error: output swing min, 141.3%. The next model corrects output swing and the large CMRR/slew mismatches.
02
Output swing correction
Model gain19.930 dBHSPICE gain20.226 dBGain error1.46%Worst error101.3%
Correct (7): DC gain, bandwidth, UGB, phase margin, ICMR max, output swing min, PSRR−.
Needs work (6): ICMR min, output swing max, both slew rates, PSRR+, CMRR.
Output swing min is fixed. Negative slew is now the largest error at 101.3%, so slew modeling becomes the main target.
03
Slew-rate refinement
Model gain19.930 dBHSPICE gain20.226 dBGain error1.46%Worst error32.9%
Full Iteration 3 comparison: prompt_stage4.txt (feedback used to create Iteration 4).
Correct (7): DC gain, bandwidth, UGB, phase margin, ICMR max, output swing min, PSRR−.
Needs work (6): ICMR min, output swing max, both slew rates, PSRR+, CMRR.
The gain model is already accurate. This revision targets slew: modeled negative slew is 7.723 V/µs versus 11.510 V/µs in HSPICE, so its error falls from 101.3% to 32.9%. The same six specs still miss the limit.
04
Negative slew converges
Model gain19.930 dBHSPICE gain20.226 dBGain error1.46%Worst error21.9%
Correct (8): DC gain, bandwidth, UGB, phase margin, ICMR max, output swing min, negative slew, PSRR−.
Negative slew passes. Positive slew is the largest remaining error at 21.9%.
05
ICMR min converges; CMRR regresses
Model gain19.930 dBHSPICE gain20.226 dBGain error1.46%Worst error57.6%
Correct (9): DC gain, bandwidth, UGB, phase margin, both ICMR limits, output swing min, negative slew, PSRR−.
Needs work (4): output swing max, positive slew, PSRR+, CMRR.
ICMR min passes, but the CMRR change overshoots to 57.6% error. The next iteration repairs CMRR without losing the nine correct specs.
06
CMRR recovery
Model gain19.930 dBHSPICE gain20.226 dBGain error1.46%Worst error17.7%
Correct (10): DC gain, bandwidth, UGB, phase margin, both ICMR limits, output swing min, negative slew, PSRR+, PSRR−.
Needs work (3): output swing max, positive slew, CMRR.
CMRR error drops to 10.4% and PSRR+ passes. Positive slew remains worst at 17.7%.
07
Output swing max converges
Model gain19.930 dBHSPICE gain20.226 dBGain error1.46%Worst error44.7%
Correct (11): all specs except positive slew and CMRR.
Needs work (2): positive slew and CMRR.
Output swing max passes. CMRR is close at 5.6%, while positive slew regresses to 44.7%.
08
Only positive slew remains
Model gain19.930 dBHSPICE gain20.226 dBGain error1.46%Worst error21.9%
Correct (12): all specs except positive slew.
Needs work (1): positive slew, 21.9% error.
CMRR now passes. The final refinement is isolated to the positive-slew equation.
09
Converged
Model gain19.930 dBHSPICE gain20.226 dBGain error1.46%Worst error4.9%
Correct (13 of 13): DC gain, bandwidth, UGB, phase margin, both ICMR limits, both output swings, both slew rates, PSRR+, PSRR−, and CMRR.
Needs work: none.
The largest remaining difference is only 4.9% (3 dB bandwidth), below the 5% limit. The scalar model is accepted and vectorized.
Absolute relative error for each valid performance metric. The dotted red line marks the 5% convergence threshold. Select the plot to open it at full resolution.
Final result
Model vs HSPICE at convergence
Iteration 9 brings all 13 modeled specs within the 5% error limit. The complete comparison is shown below.
Iteration 9 comparison
Final analytical model and simulator measurements
Percent differences use the HSPICE magnitude as the reference.
Metric
Model
HSPICE
Difference
DC gain
19.930 dB
20.226 dB
1.46%
UGB
20.452 MHz
20.144 MHz
1.53%
Phase margin
95.439°
95.433°
0.01%
ICMR min
0.433 V
0.421 V
2.86%
ICMR max
0.823 V
0.788 V
4.37%
Output swing min
0.077 V
0.080 V
3.73%
Output swing max
0.783 V
0.796 V
1.62%
Negative slew
11.035 V/µs
11.510 V/µs
4.12%
CMRR
42.187 dB
43.546 dB
3.12%
Positive slew
12.463 V/µs
12.374 V/µs
0.72%
PSRR+
19.755 dB
20.106 dB
1.75%
PSRR−
43.517 dB
45.193 dB
3.71%
3 dB bandwidth
2.062 MHz
1.966 MHz
4.85%
Reproduce and extend
Use the snapshot as a reference, not as the live workspace
# The configured driver creates and owns 5t_ota/
bash llm_aided_modelling.sh
# Compare the fresh workspace with the preserved example
diff -ru 5t_ota_example_run 5t_ota
A byte-for-byte match is not expected: LLM responses, timestamps, simulator products, and optimization results can vary. Compare schemas, stage lineage, convergence behavior, and performance trends. A real run also requires API configuration, HSPICE on PATH, and accessible foundry models.
