ATLAS

Reference implementation of ATLAS: Adaptive TDA-guided Landscape-Aware Transistor Sizing (ICCAD '26).

ATLAS is a Bayesian optimization loop for analog transistor sizing that reads the topology of the design space before deciding where to sample. At each iteration it fits one Gaussian process per specification, scores a large candidate pool with the feasibility score

h(x) = min_s (mu_s(x) - c_s) / sigma_s(x),

keeps the most promising fraction as the near-feasible set, and builds a Mapper graph over it using h as the filter. The connected components of that graph are treated as candidate feasible regions. Their number and geometry then feed the acquisition function in three places: candidates that could merge two regions or discover a new one get an exploration bonus, candidates inside a heavily-sampled region get discounted, and local candidates are drawn around region centroids in proportion to how little each region has been visited.

Installation

pip install -r requirements.txt

You also need ngspice on the PATH (tested with 44.2) and the device models for the two open PDKs. The netlists refer to them through $PDK_ROOT:

git clone https://github.com/google/gf180mcu-pdk
git clone https://github.com/google/skywater-pdk

export PDK_ROOT=/path/to/pdks     # must contain gf180/ and sky130A/

PDK_ROOT defaults to eval_engines/pdk if the variable is unset. The expected layout is

$PDK_ROOT/gf180/ngspice/{design.ngspice, sm141064.ngspice}
$PDK_ROOT/sky130A/libs.tech/ngspice/corners/{tt,ff,ss,fs,sf}.spice

Running

A single run, stopping at the first design that meets every target:

python run_experiment.py --circuit cascode_miller_opamp_gf180 \
    --method ATLAS --budget 500 --seed 42

Repeated runs of several methods on one circuit:

python run_benchmark.py --circuit cascode_miller_opamp_gf180 \
    --methods ATLAS WEIBO SCBO TuRBO MACE CMAES Random \
    --budget 500 --n-runs 10

Each run writes config.json, iterations.jsonl (per-iteration specs and a sim_time / gp_time / tda_time / acq_time breakdown), progress.log and results.json under --log-root.

To reproduce the component ablation of Table 5, disable one mechanism at a time: --no-region-discount, --no-centroid-guidance, --no-lambda-decay. The Mapper sensitivity sweep of Table 6 uses --mapper-n-cubes, --mapper-overlap, --mapper-dbscan-eps and --top-k-frac.

Circuits

Circuit	File	PDK	Params	Specs
Two-stage Miller OTA	two_stage_opamp_gf180.yaml	GF180MCU	7	gain, UGBW, PM, Ibias, Vswing, Tsettle
Cascode Miller amp	cascode_miller_opamp_gf180.yaml	GF180MCU	10	gain, UGBW, PM, Ibias, Vswing, Tsettle
Comparator	comparator_gf180.yaml	GF180MCU	6	delay, power
LDO	ldo_sky130.yaml	SKY130	13	Vdrop, PSRR (6), PM (2)

A circuit file lists the parameter bounds, the target specifications as [low, high] ranges, and the netlist for each process corner. The shipped files pin low == high, so every run of a circuit targets the same specifications and the runs differ only by their optimizer seed. Widening a range makes run_benchmark.py draw a fresh target per run instead.

Layout

topbo/
  optimizers.py   ATLAS and the baselines; the shared BO driver
  mapper.py       Mapper graph construction and bridge/frontier classification
  acquisition.py  weighted EI, the lambda schedule, the MACE ensemble
  gp_model.py     independent Matern 5/2 ARD GP per specification
  nn_model.py     probabilistic MLP ensemble, used when GP fitting gets slow
  circuit_eval.py SPICE wrapper: parameters in, specifications out
  utils.py        LHS, normalization, feasibility, figure of merit
  logger.py       per-run structured logging
eval_engines/     ngspice runner, netlists, circuit definitions

Methods

ATLAS is the proposed method. The baselines are WEIBO (weighted EI with per-spec GPs), TuRBO (trust-region BO with Thompson sampling), SCBO (constrained TuRBO), MACE (EI/PI/UCB ensemble), CMAES, and Random. The PPAAS reinforcement-learning baseline is not included here; use the authors' own release.

Settings

Defaults match the paper: 10 cover intervals at 0.3 overlap, DBSCAN with eps = 0.3 and min_samples = 3, a candidate pool of 50,000, the Mapper graph refreshed every 3 iterations, an initial design of 3 points per parameter capped at 50, an initial local radius of 0.2 decaying by 0.7 over the budget, and the exploration weight held for the first 60% of the budget before decaying quadratically. The threshold that separates bridge, frontier and interior candidates is the median distance between Mapper nodes, so it adapts to the graph rather than being tuned per circuit.

--surrogate selects the surrogate: auto (the default) starts on the GP and falls back to the NN ensemble only once GP fitting exceeds --gp-time-limit, gp pins the GP, and nn pins the ensemble. The reported experiments stay on the GP throughout. Note that the NN path does not seed torch, so nn runs are not bit-reproducible across invocations.

Notes on this release

Two defects present while the paper experiments were run have been corrected here, so numbers produced by this code can differ from the tables in the paper:

• The threshold epsilon* was derived from Mapper node positions in raw parameter units while candidate distances were measured in the normalized space. Every candidate therefore fell on the same side of the threshold and the bridge/frontier/interior split carried no signal. Node positions are now kept in the normalized space, as Section 3.3 describes.
• The constraint curriculum, which relaxes the targets handed to the surrogate while no feasible design exists, wrote into the same dictionary used to decide feasibility, so a run could stop on a design that met the relaxed targets rather than the specified ones. The curriculum now returns a separate working copy and feasibility, the figure of merit and early stopping are always judged against the target specifications.

The GP kernel is also pinned to Matern 5/2 with ARD rather than inherited from the BoTorch default, which has since changed to RBF.

Citation

@inproceedings{oh2026atlas,
  title     = {{ATLAS}: Adaptive {TDA}-guided Landscape-Aware Transistor Sizing},
  author    = {Oh, Youngmin and Won, Jihwan and Park, Yuntae and
               Hwang, Bosun and Kim, Suwan},
  booktitle = {IEEE/ACM International Conference on Computer-Aided Design (ICCAD)},
  year      = {2026},
  doi       = {10.1145/3831252.3834039}
}