Analog-DB: An Agent-First Analog Integrated Circuit Database, From Blocks to Systems
Danial Noori Zadeh, Mohamed B. Elamien
cs.AI · cs.AR · eess.SP
Quick verdict
Open, schema-governed analog design corpus re-simulating on open PDKs; credible and unusually honest, but agent claims rest on two supervised case studies.
Summary
The paper presents analog-db, an open-source, versioned database of analog integrated circuit designs released as a single schema-governed object: a process-neutral topology in a domain-specific language, reusable class-owned testbench templates, a machine-readable datasheet with per-metric specification bands, named parameters carrying matching constraints, and per-kit bindings for three open PDKs (IHP 130 nm, GF180MCU, SKY130), all simulated with ngspice. The corpus holds 68 verifiable circuits across 16 classes with 457 recorded design points; a tiered harness (T0 schema validity through T4 spec-band passing) grades each entry, derived decks regenerate byte-identically, and an independent verifier reproduced sixteen LDO baselines bit for bit. All 23 LDO circuit-kit bindings pass their own authored spec bands and 10 of 23 meet a common class band; 17 of 23 imported sizings failed and closed under a per-role gm/ID sizing loop, typically in 1-3 iterations. In two author-supervised sessions, a coding agent (Claude Opus 4.8) sized a chopper instrumentation amplifier's op-amp cores, located four hand-entry defects and a missing common-mode feedback loop that a 49-trial Bayesian sizing baseline did not repair. Main limitations, largely self-declared: agent evidence is two supervised case studies without released transcripts or ablation; structural annotation covers only 51% of devices (no CMFB template); mismatch is unmodeled so CMRR/PSRR figures are systematic upper bounds; results are typical-corner only.
Review complete
Assessment at a glance
4
risk flags identified
● high 0
● medium 2
● low 2
5 strengths · 5 weaknesses · 5 author questions
AI text signal = Binoculars zero-shot detector score over 20 sampled passages (Fast-DetectGPT co-reported) — absence does not prove human authorship, and a flag does not imply misconduct.
Levels reflect verification complexity — not a quality score. Evolve never rates your paper.
Key claims under review
- A single schema-governed object combining a process-neutral topology, class-owned testbench templates, a machine-readable datasheet with specification bands, and per-kit bindings lets analog designs be released in full and re-simulated on open PDKs.
- All 23 LDO circuit-kit bindings pass their per-entry specification bands and 10 of 23 meet a common class band; 17 of 23 imported sizings closed under a per-role gm/ID loop, typically within 1-3 iterations.
- In two supervised case studies, a coding agent working only from released artifacts sized a chopper instrumentation amplifier's op-amp cores, finding four hand-entry defects and a missing common-mode feedback loop that a 49-trial Bayesian sizing baseline did not repair.
- Byte-identical deck regeneration in CI and an independent bit-for-bit audit of sixteen LDO baselines make every reported number re-runnable from the pinned release.
- Loop phase margin alone is insufficient as a stability certificate: one corpus entry sustains a 1.84 Vpp self-oscillation while its outer-loop margin reads +69 degrees, so a zero-input transient is adopted as the deciding check.
Strengths
Reproducibility infrastructure of unusual rigor
Tiered T0-T4 verification harness with byte-identical deck regeneration gated in CI (Sec. III-C); every Appendix A table value re-simulated from the database rather than transcribed; an independent verifier reproduced sixteen LDO baselines bit for bit (Sec. IV-B); all results pinned to release v1.1.2 with ngspice and containerized open PDKs, so the flow needs no commercial license.
Honest failure and provenance reporting
Two unstable entries are withheld by name (011_peng_iac, 021_yan_az); optimizer artifacts are flagged rather than buried (Table X # rows); the 030 composite is recorded below the 45-degree PM floor at 33.3 degrees on ihp-sg13g2; and the agent-vs-optimizer comparison is explicitly disclaimed as asymmetric and non-controlled (Sec. IV-D).
Self-audit of its own measurement apparatus
Genuinely machine-actionable contract
Transferable methodological finding on stability
Weaknesses
Agent evidence is anecdotal and unarchived
Validated status is partly self-fulfilling
Structural annotation is half-complete
Simulation realism ceiling
Corpus skew and thin classes
Full Review Report
$1.99 · one-timeUnlock the complete preliminary analysis Evolve produced for this paper.
Report preview
Scope & Claims
This is a resource/benchmark paper introducing analog-db (release v1.1.2, github.com/MacAnalog/spicexplorer-release): a versioned corpus of 68 verifiable analog circuits across 16 classes, each released as one schema-governed object E=(T,Θ,C,B) - a process-neutral topology, a parameter space with tie groups/ratios/freezes, a machine-readable datasheet contract, and per-kit bindings for ihp-sg13g2, gf180mcu and sky130, all simulated in ngspice. Central claims: (1) the representation removes both sharing barriers the authors identify (NDA encumbrance; missing testbenches); (2) all 23 LDO circuit-kit bindings pass their own recorded spec bands, 10 of 23 meet a common class band, and 17 of 23 imported sizings closed under a per-role gm/ID loop, typically in 1-3 iterations (Sec. IV-D); (3) in two supervised sessions a coding agent (Claude Opus 4.8) repaired and sized a chopper instrumentation amplifier from released artifacts alone, finding four hand-entry defects and a missing CMFB loop that a Bayesian sizing-only baseline missed (Sec. IV-D, Tables V-VI). Claims are unusually well hedged for this space.
The remaining sections cover Methodology Assessment, Statistical Evidence, Results & Reproducibility, Limitations & Risks, and Verdict & Recommendations.
- All 5 strengths & 5 weaknesses, analyzed in detail
- 4 risk flags with severity, section & reviewer questions
- Methods & statistical evidence assessment
- 5 questions a peer reviewer would ask
- Full six-dimension evaluation report & recommendations
One-time purchase · this paper only
One-time purchase, this paper only. AI-generated preliminary analysis — not peer review.
Frequently asked questions
What is analog-db and what does it contain?
analog-db is an open-source, versioned database of analog integrated circuit designs (68 circuits across 16 classes as of release v1.1.2). Each entry packages a process-neutral circuit topology, reusable testbench templates, a machine-readable datasheet with pass/fail specification bands, named parameters with matching constraints, and bindings to three open-source process design kits (IHP 130 nm, GF180MCU, SKY130). Everything simulates with the free ngspice simulator, so anyone can re-run the published results without commercial EDA licenses.
Is the analog-db code and data publicly available?
Yes. The paper releases everything at github.com/MacAnalog/spicexplorer-release, release v1.1.2, which is the archived revision all reported numbers come from. The authors state that all benchmark-table values were re-simulated from the database rather than copied from source papers, derived simulation decks regenerate byte-identically under continuous integration, and an independent verifier reproduced sixteen regulator baselines bit for bit.
Did an AI agent really design an analog circuit in this paper?
Not autonomously. Two author-supervised case studies show a coding agent (identified as Claude Opus 4.8) sizing a low-dropout regulator and the op-amp cores of a chopper instrumentation amplifier until they pass their datasheets, diagnosing defects such as a missing common-mode feedback loop. The authors set objectives and reviewed every edit, transcripts are not released, and the paper itself labels the evidence as feasibility under supervision, not autonomy.
How does analog-db compare to AnalogGym?
AnalogGym ships 30 analog topologies with only two shared testbench templates (amplifier and LDO), no per-circuit acceptance bands, no recorded design points, and flat non-composable netlists. analog-db releases 68 circuits where 65 carry recorded, spec-judged design points, verified blocks compose into larger systems, and a schema-governed file contract lets AI agents query and reuse entries directly without a server (Table I of the paper).
About this review
This is an AI-generated preliminary review by Evolve.Pub, produced from the openly available text of the paper. It is not peer review and does not represent the authors' views. The canonical record of the paper is at the publisher link above.
- The provided text is a sampled excerpt with marked middle gaps (parts of Secs. II-III-A, portions of Sec. IV-A, and the tail of Sec. V-A through V-B); judgments touching those regions rely on surrounding text and the abstract.
- Table IV is cut mid-row in the excerpt (the LDO row truncates), so per-class verification counts beyond amplifiers could not be checked against the text.
- The task's abstract metadata is cut mid-sentence, but the embedded full text contains the complete abstract; no content claim here relies on the metadata version.
Go deeper on this paper
Wrote a paper? Stress-test it before reviewers do.
Upload your manuscript for the same structured scrutiny — reasoning, statistics, evidence, figures, and reproducibility.