Wire-aware component placement for generated schematics

1 Introduction

Component placement and wire routing are coupled decisions. Moving one symbol changes route length, bend locations, local density, and the space available to neighboring nets. A sequential system can lock in poor geometry before routing begins.

Schematic placement is also different from physical board placement. The objective is communication, not only physical feasibility. Engineers use position, orientation, and grouping to show circuit structure.

We study a wire-aware placement method for generated schematics. The method creates several positions for one missing component. It evaluates each position with explicit geometric and wiring terms.

The study uses reconstruction because it provides a human reference position. The reference supports measurement but does not control production scoring. Validity remains independent of similarity.

2 Placement task

A task begins with a parsed schematic and one selected component. The extractor records the component position, orientation, size, pins, and connected anchors. It then removes the component from a copy of the scene.

The remaining symbols and text become placement obstacles. The original sheet bounds and grid define the legal search space. The original connectivity graph remains the expected electrical context.

Each candidate contains a point, orientation, bounding box, transformed pin positions, and estimated connections. The candidate also records its generation motif. This provenance supports later error analysis.

3 Candidate generation

The generator uses several placement motifs. Region-center and block-grid motifs provide regular structure. Anchor-biased and net-centroid motifs use electrical context.

Component-proximity and type-affinity motifs preserve common visual relationships. Passive motifs place small support parts near their likely owner. Rail-terminal motifs handle power relationships at block boundaries.

The generator evaluates four orthogonal orientations when the component permits them. It snaps candidate points to the schematic grid. Duplicate positions are removed before scoring.

Candidate diversity is necessary because no single motif works for every component role. A connector, decoupling capacitor, processor, and rail symbol have different placement patterns. The scorer compares these alternatives under one explicit objective.

4 Objective function

The objective contains four groups of terms. Spatial terms measure region position, local density, grid state, and sheet bounds. Electrical terms measure pin distance and distance to the connected-net centroid.

Relational terms measure distance to connected components and affinity with the component role. Routing terms estimate whether the transformed pins can reach their anchors. They also estimate route length and bend cost.

Collision, off-grid, and out-of-bounds states receive dominant penalties. The selector also sorts by validation state. A strong preference score cannot promote an invalid placement.

5 Route-feasibility probe

A wire-aware score must use transformed pin geometry. The placement engine computes each pin position after translation and rotation. It then forms provisional connections to known anchors.

A limited routing probe evaluates the most promising placement candidates. The probe returns validity, route length, bend count, and selected route identity. Invalid or difficult probes add cost to the placement.

Limiting the probe to top candidates controls runtime. Cheap spatial features remove clearly poor placements first. Detailed route evaluation is reserved for plausible positions.

This order is a coarse-to-fine search. It links placement to downstream wiring without running a complete sheet router for every point. The approximation is measurable through the recorded route terms.

6 Evaluation protocol

The evaluation uses component-removal tasks from parsed public schematics. Each task preserves the scene around one removed component. The human point and orientation remain available for similarity measurement.

The main measures are selected pass rate, valid candidate rate, and mean human similarity. Selected pass rate tests the validity-first policy. Valid candidate rate measures search difficulty.

Human similarity measures geometric agreement with the reference placement. It does not prove that the reference is the only good position. A lower similarity can still be valid and readable.

The recorded evaluation contains 20 tasks and 888 candidates. The mean pool size is 44.4 candidates per task. All tasks are evaluated after candidate validation.

7 Results and discussion

The valid candidate rate was 60.7 percent. This value shows that the generator explored many positions that violated at least one hard constraint. A preference-only selector would face substantial risk.

The selected pass rate was 100 percent. Every evaluated task contained at least one valid option, and the selector chose a valid option each time. This result supports the lexicographic validity rule.

Mean human similarity was 0.772. The score indicates substantial positional agreement without exact copying. Route-aware and relational terms can favor a nearby alternative when it offers easier wiring.

The evaluation does not isolate the effect of each cost group. An ablation study should remove routing, net-centroid, and motif terms in separate runs. That study would quantify each contribution.

8 Limitations

Twenty tasks are not sufficient for broad statistical inference. The result demonstrates pipeline behavior, not production generalization. Larger held-out sets are required.

Reconstruction also gives the system more context than new design synthesis. Neighbor positions and connected anchors already exist. A new schematic must create that structure jointly.

Human similarity can reward source-specific habits. Future evaluation should include blinded engineer preference and downstream routing quality. These measures can identify valid improvements that differ from the reference.

The route probe is approximate. A placement that appears easy in isolation can create conflict during full-sheet routing. Joint multi-component optimization remains future work.