Skip to content

Optimizers API

OptimizationProperty

Defines one optimization variable.

Fields include:

  • name
  • type (MODEL_INPUT or NETLIST_VARIABLE)
  • min_value
  • max_value
  • step
  • optional unit
  • optional linked_to
  • optional initial_value: the value the optimizer evaluates first

OptimizationType

  • MODEL_INPUT: variable passed to ONNX surrogate inputs.
  • NETLIST_VARIABLE: variable patched directly in parsed netlist content.

DesignGoal and DesignParameter

DesignGoal combines:

  • parameter enum (DesignParameter),
  • optional min/max bounds,
  • optional frequency_range,
  • optional weight,
  • optional scale, the unit a miss of a bound is measured in.

Penalty is the mean over the goal's points of the squared bound violations, each divided by scale; without one, by the bound (|bound|, or 1 for a bound of 0; see deviation_scale). The mean keeps a band goal's share independent of its point count.

Each DesignParameter declares the analysis it needs, so goals are grouped by analysis and evaluated against the matching result. with_variant(parameter, variant) reads a parameter from an AnalysisVariant (the analysis run on the netlist with some values changed) instead, under the name <variant>:<name>. make_expression_parameter(name, expression, terms) builds an ExpressionParameter: a formula such as "a - b" over named terms (other parameters, possibly expressions themselves, possibly from different analyses), read at the goal's frequency and parsed by GoalExpression from a restricted grammar. Small-signal parameters (S-parameters, Lp, Qs, k, mu, SRF, …) read the .AC network; the spectrum parameters read the HB or transient spectrum:

Factory Parameter name Description
make_power_dbm(node, simulation_type) Power_dBm[<node>] Output power in dBm at an analysis point
make_gain_db(port, amplitude, z0, node, simulation_type) Gain_dB[<port>@<node>] Transducer gain in dB referred to a port's drive level
make_isolation_db(node, simulation_type) Isolation_dB[<node>] Margin in dB between the target line and the strongest other line

simulation_type defaults to SimulationType.HB; passing SimulationType.TRAN evaluates the same quantity on the transient spectrum and prefixes the name with TRAN:. These are built per netlist rather than taken from the static catalogue, because the available nodes and ports differ per circuit. See Advanced -> Harmonic Balance and Advanced -> Transient Analysis.

OptunaOptimizer

Default optimizer wrapper for Optuna.

Typical options:

  • sampler: tpe, random, simulated_annealing, auto. TPE uses Optuna's defaults: multivariate for a single objective, independent for several; override them through sampler_kwargs, e.g. {"multivariate": False}
  • pruner: median, successive_halving, hyperband
  • multi_objective: boolean

A parameter's initial_value makes the first trial evaluate it; the others are sampled.

GradientDescentOptimizer

Used for local refinement workflows, including optional fine-tuning mode in COBRA. It is simultaneous-perturbation (SPSA) gradient descent in coordinates normalised to each parameter's range: every step probes the current point at ± a small distance along a random direction (two trials), then moves against the estimated gradient. Both distances shrink with every step. The point is kept unrounded, so steps below a parameter's grid still add up; only the proposals are rounded to it.

  • learning_rate: first step as a fraction of the range (default 0.2)
  • exploration_scale: probe distance; 1 probes at 10 % of the range, at least one grid step (default 0.5)
  • random_seed

It starts at the context's values (in EM fine-tuning, the design just evaluated), else at each parameter's initial_value, else in the middle of the range; a start not taken from the context is evaluated first. It needs the result of one trial before it suggests the next, so parallel_trials must be 1.

Tip

Start with OptunaOptimizer and move to gradient descent only when you are refining near a known good region.