betaFeedback

WorksComputer Science

NEMESIS2.0: LLM-Aided Netlist-Driven Performance Equation Modeling for Accelerated Analog Circuit Sizing

Abstract

This work presents NEMESIS2.0, an open-source LLM-aided framework that automatically derives equation-based performance models for analog circuits. The resulting equations provide fast, accurate surrogates for SPICE evaluation. To validate the method, equation-based models are developed for seven circuit test cases in a commercial 65nm CMOS technology, covering 15 metrics on five operational transconductance amplifier topologies, and six metrics on two dynamic comparator topologies. When tested on 1000 unseen variants of these seven designs, each corresponding to a different set of independently sampled transistor sizes, the accuracy is found to be within 5% of SPICE. To demonstrate use cases for NEMESIS2.0, the generated models are used as performance evaluators in multiple transistor sizers based on different population-based stochastic search optimizers. For each of the seven test cases, transistor sizing is performed for 100 different sets of performance specifications, with final SPICE simulation confirming that 83.3% of the resulting designs meet all specifications after a single sizing run. The rest fail to meet the specifications due to discrepancies between the model and SPICE, and require iterative resizing under tightened specifications; all such scenarios meet specifications within five iterations. Within the transistor sizers, this equation-based performance analysis evaluates ~400 design candidates per second, versus 0.14 candidates per second for SPICE-based evaluation, achieving 196-238x sizing runtime speedup, including final SPICE verification.

Introduction

Analog integrated circuits interface physical signals with digital systems through sensing, amplification, and data conversion. Their performance depends on circuit topology, device dimensions, operating regions, parasitics, and loading conditions, which collectively influence multiple, often competing, performance metrics. Meeting the desired specifications therefore requires systematic optimization of circuit parameters to navigate these performance tradeoffs. For a selected topology, schematic-level sizing determines transistor dimensions and bias conditions to meet the design specifications. To automate this process, optimization algorithms iteratively propose candidate transistor sizes and use a performance evaluator to estimate performance metrics of the design. These estimates guide candidate selection and refinement until the objectives are met or the process times out. The evaluator is therefore central to automated sizing: slow evaluation limits the number of candidates that can be explored within a fixed runtime budget, while inaccurate predictions can misguide the optimizer.

Several automated sizing approaches employ SPICE-based evaluation, including methods based on stochastic search [1, 2, 3, 4, 5, 6, 7, 8], Bayesian optimization [9, 10, 11], reinforcement learning [12, 13, 14], and large language models (LLMs) [15, 16]. SPICE provides accurate performance estimates, but it is computationally expensive, particularly over the large number of candidate sizes that are evaluated by sizing algorithms. To reduce this runtime cost, data-driven surrogates, trained on SPICE-generated samples, have been used to estimate circuit performance using response-surface models [17, 18], probabilistic models [9, 10, 11], neural networks [19, 20], and graph-based models [21]. Although they are faster than SPICE, these surrogates require extensive training data and adequate design-space coverage to achieve accurate predictions. Moreover, while SPICE and data-driven surrogates provide performance estimates, they do not inherently expose the analytical relationships between device parameters and circuit performance metrics. Such relationships can provide additional interpretability, allowing circuit designers to understand how individual device parameters influence performance, identify potential sources of performance degradation, and reason about design tradeoffs.

Analytical modeling offers a way to explicitly capture these relationships without learning them primarily from sampled design–performance data. Methods based on analytical equations [22, 23, 24, 25, 26, 27], symbolic analysis [4], and geometric programming [28, 29, 30] follow this principle, providing explicit performance models that can be evaluated rapidly without requiring large training datasets. However, their reliance on simplified MOS models and operating-region-specific assumptions can compromise accuracy as device dimensions, bias conditions, and transistor operating regions change during sizing. Moreover, these equations are required for a vast range of circuits, each of which may have multiple topologies and differing performance metrics: deriving and adapting equations across circuit topologies and performance metrics requires substantial manual effort. These limitations motivate automatically generated equation-based performance models that retain the speed and structured, interpretable circuit-level representation of analytical methods while achieving SPICE-referenced accuracy across varying operating conditions, thereby providing fast evaluators for schematic-level sizing.

To address these challenges, we present NEMESIS2.0, an LLM-aided framework that extends our NEMESIS work [31] to automatically generate SPICE-verified, interpretable performance equations across circuit topologies. As shown in Figure 1(a), the framework integrates circuit interpretation, operating-point selection, equation generation, and SPICE-guided refinement to construct reusable performance prediction models. We integrate these models as performance evaluators into an automated sizing framework (Figure 1(b)) where population-based optimization algorithms are used to optimize transistor dimensions and bias conditions under application-specific performance objectives and constraints. By using analytical equations for performance estimation, the framework reduces the computational cost of evaluating candidate designs without repeated metric-specific SPICE simulations, while retaining full SPICE verification of the final optimized designs.

Figure 1. (a) NEMESIS2.0 performance modeling framework and its application in (b) the automated sizing framework.

The performance models in NEMESIS2.0 are constructed through a sequential workflow primarily involving three LLM-based agents. First, the Topology Analysis Agent interprets circuit connectivity to identify transistor roles and selects representative operating-point configurations. Next, the Equation Generation Agent uses the identified circuit structure and operating-point configurations to derive analytical equations relating device-level electrical parameters to circuit performance metrics and evaluates their accuracy against SPICE simulations. Finally, the Equation Repair Agent selectively refines equations that fail to meet the prescribed SPICE-referenced error threshold.

The Topology Analysis Agent begins by analyzing the circuit netlist and schematic to identify transistor roles and functional primitives, such as differential pairs (DPs) and current mirrors (CMs). To capture variations in circuit performance across diverse transistor operating conditions, the agent selects multiple operating-point configurations using the transconductance-efficiency ratio (gm/IDg_m/I_D), considering the functional role of each transistor. The gm/IDg_m/I_D ratio governs transistor inversion levels and influences tradeoffs among speed, intrinsic gain, power efficiency, and other performance characteristics [32, 33, 34]. Accordingly, the agent constructs three representative configurations targeting high-speed, balanced, and high-gain operation, with transistor operating regions selected according to their functional roles. These configurations provide diverse operating conditions for generating and verifying analytical equations, improving their robustness.

Using the circuit interpretation and operating-point configurations provided by the Topology Analysis Agent, the Equation Generation Agent constructs an analytical equation for each performance metric in terms of device-level small- and large-signal parameters, such as transconductance (gmg_m), output conductance (gdsg_{ds}), drain current (IDI_D), and parasitic capacitances. Retrieval-augmented generation (RAG) [35] supports equation construction by retrieving previously verified performance equations for similar circuits. To predict circuit performance using the generated equations, the required device parameters must be obtained. Accordingly, the Topology Analysis Agent maps the assigned gm/IDg_m/I_D targets to realizable transistor dimensions using technology-characterized lookup tables (LUTs). Rather than relying on simplified MOS device models [25, 26, 27], NEMESIS2.0 extracts the device parameters through computationally inexpensive SPICE operating-point (.OP) analyses [31]. The extracted parameters are then substituted into the generated equations to estimate circuit performance across all three operating-point configurations.

To assess the accuracy of the generated equations, their predicted performance metrics are compared against reference values obtained from metric-specific SPICE simulations across all three operating-point configurations. Equations that meet the prescribed error threshold are retained, while those that exceed the threshold are forwarded to the Equation Repair Agent for refinement. Guided by the verification feedback, the agent adds or removes relevant small- and large-signal parameters based on LLM reasoning to improve equation accuracy while preserving the underlying circuit relationships. The verification and refinement process continues until the equations satisfy the error threshold across all three configurations, yielding one SPICE-verified analytical equation per performance metric.

The proposed modeling workflow introduces two key extensions over our earlier NEMESIS framework [31]. First, whereas NEMESIS generated performance equations around a single application-specific operating-point configuration, NEMESIS2.0 constructs and verifies equations across three representative configurations spanning different performance tradeoffs. This broader coverage improves equation robustness across diverse transistor operating conditions. Second, while NEMESIS employed a single agent for equation generation and repair, NEMESIS2.0 separates these responsibilities into dedicated agents, enabling targeted refinement based on SPICE verification feedback. Together, these extensions improve the generality of the performance equations and provide a more structured approach to their generation and refinement.

To demonstrate the applicability of the generated equations to automated circuit sizing, we construct performance models for five operational transconductance amplifier (OTA) and two dynamic comparator (COMP) topologies and integrate them into a population-based sizing framework. As a post-processing task, an LLM-based Vectorization Agent converts the verified analytical equations into executable vectorized models, enabling efficient performance prediction for multiple candidate sizing configurations. To assess their portability across optimization methods, we independently employ three algorithms: genetic algorithm (GA) [36], particle swarm optimization (PSO) [37], and differential evolution (DE) [38]. Each algorithm optimizes transistor dimensions and bias conditions for a fixed circuit topology under the same performance objectives and constraints, using the equation-based evaluators. Finally, the optimized designs are verified against the original target specifications through SPICE simulations.

The key contributions of this work are:

  • Robust performance models: Analytical equations relate

transistor electrical parameters to up to 15 circuit performance metrics, meeting a 5% SPICE-referenced error threshold capturing speed, gain, and power tradeoffs.

  • Automated modeling across circuit topologies: Dedicated LLM-based agents automate performance modeling and targeted refinement, demonstrated on five OTA and two COMP topologies without manual equation derivation.

  • High-throughput performance evaluation: SPICE-verified equations enable accurate performance estimation for ≈400\approx400 candidate designs per second, avoiding repeated metric-specific SPICE simulations during optimization.

  • Specification-compliant automated sizing: Across seven circuits, each sized for 100 different specification sets, 83.3% of designs meet all targets on the first attempt in <35<35sec; the rest succeed within five iterative sizing attempts.

  • Accelerated automated sizing: By avoiding repeated metric-specific SPICE simulations during optimization, the framework achieves 196×196\times–238×238\times speedups over SPICE-in-the-loop sizing for single-attempt cases and 39×39\times–50×50\times for five-attempt cases, while retaining final SPICE verification.

The remainder of this paper is organized as follows. Section II presents the NEMESIS2.0 performance-modeling framework, Section III describes its application in the agentic sizing framework, and Section IV presents the experimental results. Section V concludes the paper.

NEMESIS2.0 Modeling Framework

Figure 2 illustrates the complete modeling workflow, from circuit interpretation to the generation of SPICE-verified performance equations. The flow begins with the Topology Analysis Agent, which processes the circuit netlist and schematic to identify primitives, transistor roles, and signal paths, and establishes three operating-point profiles with assigned gm/IDg_m/I_D values, branch currents, and channel lengths. Guided by this circuit interpretation, the operating-point profiles, and relevant knowledge retrieved from the RAG knowledge base, the Equation Generation Agent constructs analytical equations and their executable implementations for each performance metric.

Figure 2. Overview of the NEMESIS2.0 performance modeling framework.

For each profile, the operating-point anchoring stage uses technology-characterized LUTs to map the assigned gm/IDg_m/I_D

values and channel lengths to transistor current densities, from which transistor widths are determined using the branch currents. SPICE .OP analyses then provide the device parameters needed to evaluate the metric equations, while metric-specific DC, AC, and transient simulations establish reference performance values. The verification stage compares equation predictions against these SPICE references across all three profiles, retaining equations that satisfy the prescribed error threshold and forwarding inaccurate equations to the Equation Repair Agent. The agent iteratively refines the equations while reusing previously extracted device parameters and SPICE reference values, avoiding redundant simulations. Once verified, the equations and their executable implementations are stored in the RAG knowledge base for future reference. The following subsections describe each stage of the flow in detail, using a two-stage OTA (2S-OTA) (Figure 3(a)) comprising a five-transistor OTA (5T-OTA) (Figure 3(b)) first stage and a common-source (CS) second stage as a running example.

Figure 3. (a) Schematic of the 2S-OTA; the dashed boundary encloses the first stage, corresponding to (b) the previously modeled 5T-OTA from the RAG knowledge base.

Multimodal circuit interpretation and knowledge retrieval

The flow begins with the multimodal-input block in Figure 2, where the transistor-level netlist and schematic are jointly analyzed to identify circuit structures and device roles. The netlist specifies device connectivity, terminal assignments, model types, and parameters, while the schematic provides visual cues for symmetry, multiple stages, and signal flow that may be less apparent from connectivity alone. Together, these inputs enable the Topology Analysis Agent to identify circuit topology, functional primitives, and device roles, and construct operating-point profiles. The resulting circuit interpretation also guides the Equation Generation Agent in identifying the physical mechanisms relevant to each performance metric. The significance of incorporating schematic information alongside the netlist is further examined in Section IV-D.

To complement circuit interpretation with knowledge from previously modeled circuits, we represent the transistor-level netlist as a terminal-labeled bipartite graph, G=(Vdev∪Vnet,E,ℓv,ℓe)G = (V_{\mathrm{dev}} \cup V_{\mathrm{net}}, \mathcal{E}, \ell_v, \ell_e), where VdevV_{\mathrm{dev}} and VnetV_{\mathrm{net}} denote device and net vertices, and E\mathcal{E} denotes the edges connecting device terminals to nets respectively. The vertex labels ℓv\ell_v distinguish device types and identify nets connected to external circuit pins, while the edge labels ℓe\ell_e preserve device-terminal roles such as drain, gate, source, and bulk. External circuit pins retain functional labels identifying signal inputs, outputs, power supplies, and bias terminals. In contrast, transistor instance names and internal net names are excluded because they do not affect circuit connectivity or device-terminal relationships. Device dimensions are also omitted to represent circuit structure independently of a particular sizing configuration. The resulting graph therefore preserves terminal-aware connectivity and external pin functionality while remaining insensitive to changes in device naming, internal net naming, and transistor sizing. The RAG knowledge base stores these representations alongside previously modeled circuit information and SPICE-verified performance equations.

To retrieve structurally related circuits, we evaluate similarity using two complementary descriptors: a terminal-aware Weisfeiler–Lehman (WL) fingerprint [39] capturing local device-to-net connectivity and an analog-motif vector capturing functional structures. The latter records device composition, diode-connected devices, shared-gate groups, differential-pair and current-mirror patterns, and output, supply, and bias connections. For a query circuit qq and stored circuit ii, the similarity score is S(q,i)=αsWL(q,i)+(1−α)smotif(q,i)S(q, i) = \alpha s_{\mathrm{WL}}(q, i) + (1 - \alpha)s_{\mathrm{motif}}(q, i), where sWLs_{\mathrm{WL}} is the multiset-Jaccard similarity [40] between WL fingerprints and smotifs_{\mathrm{motif}} is the cosine similarity between motif vectors. We heuristically set α=0.6\alpha= 0.6 for all queries, assigning slightly greater weight to terminal-level connectivity while retaining information about higher-level circuit structures. Stored circuits are ranked according to S(q,i)S(q, i), and the top NRet=3N_{\mathrm{Ret}} = 3 records are supplied to the Equation Generation Agent.

Figure 3 illustrates this retrieval process using the 2S-OTA and the previously modeled 5T-OTA. The colored blocks highlight the circuit primitives identified by our approach and are not part of the input. The shared 5T-OTA first stage provides functional similarities captured by the analog-motif descriptor, whereas the additional CS stage and compensation network alter the connectivity and reduce the WL similarity. The combined similarity score therefore retrieves the 5T-OTA as a related, but nonidentical, circuit. Its stored structural information and SPICE-verified equations provide references for modeling the shared first stage. Using these references together with the target circuit interpretation, the Equation Generation Agent accounts for the additional CS stage, compensation elements (RCR_{\mathrm{C}} and CCC_{\mathrm{C}}), and modified signal and loading paths in the 2S-OTA. Thus, retrieved equations guide the analytical derivation without being transferred directly to the target circuit. We independently verify the generated equations against SPICE for the 2S-OTA, and only verified equations are added to the RAG knowledge base.

Multi-region gm/IDg_m/I_D operating-point profile generation

Building on the circuit interpretation and retrieved knowledge, the Topology Analysis Agent constructs three representative operating-point profiles to capture variations in circuit behavior across transistor inversion regions. Changes in inversion region affect device-level electrical characteristics and their contributions to circuit performance, making it important to account for these variations when developing analytical models. The agent generates high-speed, balanced, and high-gain profiles by assigning different combinations of strong, moderate, and weak inversion to functional transistor groups. These profiles provide distinct operating conditions for evaluating the analytical equations without exhaustively exploring all possible operating-region combinations.

To construct these profiles, the agent first groups transistors according to their functional roles and shared operating-point constraints rather than circuit stages alone. The grouping is derived from the supplied netlist, schematic, and prompt instructions, with additional guidance from structurally related circuits retrieved through the RAG mechanism described in § 2-A. Retrieved records, when available, provide reference primitive classifications, connectivity relationships, and gm/IDg_m/I_D assignments; otherwise, the agent determines these from the target circuit information. The left column of Figure 4 illustrates the resulting grouping for the 2S-OTA: G1 comprises the input differential pair, G2 the PMOS active-load current mirror, G3 the NMOS bias-derived current-mirror devices, and G4 the second-stage gain device. Although M6 belongs to the second stage, it is grouped with M4 and M5 in G3 because these devices share a bias-derived gm/IDg_m/I_D constraint. Within these groups, matched devices receive equal gm/IDg_m/I_D targets, while current-mirror devices additionally share channel lengths, with their branch currents determined by the corresponding mirror ratios.

Topology analysis agent output of 2S-OTA showing topology interpretation and assigned operating-point profiles

Figure 4. Topology analysis agent output of 2S-OTA showing topology interpretation and assigned operating-point profiles.

To assign numerical gm/IDg_m/I_D targets, we provide the agent with inversion-region ranges and primitive-specific sizing guidelines derived from established gm/IDg_m/I_D methodologies [32], [34]. The agent selects values appropriate to the target technology, using assignments from retrieved circuits as additional references when available. The right column of Figure 4 shows the resulting profiles for the 2S-OTA. In the high-speed profile, G1 and G2 use gm/ID=9g_m/I_D = 9, and G4 uses gm/ID=6g_m/I_D = 6, placing these groups in strong inversion, while G3 operates in moderate inversion at gm/ID=11g_m/I_D = 11. The balanced profile places G1–G3 in moderate inversion while retaining G4 in strong inversion for output drive. In the high-gain profile, G1 shifts to weak inversion at gm/ID=18g_m/I_D = 18, while G2–G4 operate at higher gm/IDg_m/I_D values within moderate inversion. These assignments vary device operating conditions while preserving the circuit’s functional and biasing relationships. The resulting gm/IDg_m/I_D targets, branch currents, and channel lengths are subsequently mapped to transistor dimensions using technology-characterized lookup tables (LUTs). SPICE operating-point analyses of the resulting profiles then provide distinct sets of device parameters for evaluating the same analytical equations under different operating conditions.

Performance equation generation

The Equation Generation Agent constructs topology-specific analytical performance equations using the circuit interpretation, operating-point profiles, and the retrieved knowledge from the RAG knowledge base. For each requested performance metric mm, the agent generates a single equation that is evaluated across all three operating-point profiles:

y^m,k=fm,N({xd,k}d∈D,u),(1)\hat{y}_{m,k}=f_{m,\mathcal{N}}\left(\{\mathbf{x}_{d,k}\}_{d\in\mathcal{D}},\mathbf{u}\right), \tag*{(1)}

where fm,Nf_{m,\mathcal{N}} denotes the equation for metric mm and netlist N\mathcal{N}, and y^m,k\hat{y}_{m,k} is its prediction at operating-point profile kk. The set D\mathcal{D} contains the circuit transistors, while xd,k\mathbf{x}_{d,k} represents the profile-dependent information for transistor dd. This includes device dimensions—width (WW) and channel length (LL); DC operating-point quantities—terminal currents and voltages, threshold voltage (VthV_{\mathrm{th}}), and saturation voltage (VDSATV_{\mathrm{DSAT}}); and small-signal parameters—transconductance (gmg_m), output conductance (gdsg_{ds}), body transconductance (gmbg_{mb}), and intrinsic capacitances. The vector u\mathbf{u} contains external circuit and testbench parameters, such as supply voltages and load capacitance, that remain fixed across the three profiles.

To construct these equations, we prompt the agent to identify the devices, signal paths, and circuit relationships governing each performance metric in the target topology. Retrieved equations provide references for relevant analytical relationships, but the agent is instructed to account for differences in transistor connectivity, device functions, and loading conditions before adapting them to the target circuit. To reduce the risk of hallucinations, we further instruct the agent to avoid introducing devices, nodes, or circuit parameters that are absent from the input netlist.

Example: 2S-OTA. Representative equations for the 2S-OTA illustrate how the Equation Generation Agent translates circuit interpretation into analytical performance models. For illustration, we compare the generated DC-gain equation with a conventional first-order model. Assuming matched devices and an ideal current mirror, the manual model approximates the overall gain as the product of the two stage gains [41] as:

DC_Gainmanual≈20log⁡10∣gm,M1gds,M1+gds,M3⋅gm,M7gds,M6+gds,M7∣.\mathrm{DC\_Gain}^{\mathrm{manual}}\approx20\log_{10}\left\lvert\frac{g_{m,M1}}{g_{ds,M1}+g_{ds,M3}}\cdot\frac{g_{m,M7}}{g_{ds,M6}+g_{ds,M7}}\right\rvert.

Using the identified differential pair (M0–M1), PMOS current mirror (M2–M3), and second-stage common-source amplifier (M7), the agent instead generates:

DC_Gain: 20*log10(abs(0.5*(gm_M1+gm_M0)*(gm_M3)/(gm_M2+gds_M2))/(gds_M1+gds_M3)*(gm_M7)/(gds_M6+gds_M7))) Reasoning: Differential gain current from M1 and the PMOS current mirror through M2-M3 drive the first-stage output resistance, followed by the gain of the M7 stage.

The generated equation separately captures the direct contribution of M1 and the contribution of M0 transferred through the CM, as indicated by the agent reasoning. The factor gm,M3/(gm,M2+gds,M2)g_{m,M3}/(g_{m,M2}+g_{ds,M2}) accounts for unequal mirror gmg_m and the finite conductance of the diode-connected M2. Under matched-device and ideal-mirror assumptions, the expression reduces to the conventional model. This comparison illustrates how the agent incorporates physically meaningful device-level effects without manual derivation.

The 3-dB-bandwidth (BW_3dB) equation is derived from the dominant pole at the first-stage output, accounting for the drain capacitances of M1 and M3, the gate capacitance of M7, and the Miller-multiplied compensation capacitance:

BW_3dB: (gds_M1+gds_M3)/(2*pi*(cd_M1+cdd_M3+cgg_M7+C*(1+abs(gm_M7)/(gds_M6+gds_M7)))) Reasoning: 1st-stage dominant pole includes M7 gate capacitance, M1/M3 drain capacitances, and Miller-multiplied Cc through the 2nd-stage gain.

For the output common-mode range (OCMR), the agent uses the saturation constraints of the output-stage NMOS device M6 and PMOS device M7 to determine the lower and upper output-voltage limits:

OCMR_Min/Max: [VSS+abs(vdsat_M6), VDD-abs(vdsat_M7)] Reasoning: M6 and M7 saturation boundaries determine the minimum and maximum output voltages, respectively.

This 2S-OTA example illustrates how the agent constructs initial analytical equations for different performance metrics from the identified circuit structure, with the remaining metrics modeled similarly. Each equation is accompanied by a Python-executable implementation that accepts device and testbench parameters to estimate performance. At this stage, the equations remain unverified and are subsequently evaluated against SPICE and selectively repaired, as described in § 2-D.

SPICE-anchored equation verification and repair

SPICE-based equation verification

Although the generated equations capture physically motivated circuit relationships, they may omit relevant device contributions, loading effects, or parasitic capacitances, leading to inaccuracies across operating conditions. As in the original NEMESIS framework [31], we assess their accuracy by comparing equation predictions against reference values obtained from metric-specific DC, AC, and transient SPICE simulations. In NEMESIS2.0, we extend this verification across the three gm/IDg_m/I_D operating-point profiles—high-speed (OP1), balanced (OP2), and high-gain (OP3)—by evaluating the same analytical equation using the device parameters corresponding to each profile. The metric-specific testbenches are described in Section IV.

To reduce inaccuracies associated with simplified transistor models, we evaluate the analytical equations using device parameters extracted using .OP analysis from SPICE at the corresponding operating points. Consider, for example, a resistively loaded NMOS common-source stage, for which a candidate low-frequency voltage-gain (AvA_v) relationship is

Representative inputs to the Equation Repair Agent

Figure 5. Representative inputs to the Equation Repair Agent: (left) metric- and topology-agnostic repair instructions; (right-top) representative actual SPICE and model values supplied for a failed metric; and (right-bottom) signed model vs. SPICE error summary across the operating profiles for all the metrics.

Modeling UGF of the 2S-OTA testcase

Figure 6. Modeling UGF of the 2S-OTA testcase.

Av≈−gmgout,gout=gds+1RL,(2)A_v \approx-\frac{g_m}{g_{out}}, \qquad g_{out}=g_{ds}+\frac{1}{R_L}, \tag*{(2)}

where RLR_L is the load resistance. Conventional first-order MOS analysis may approximate the device parameters as

gm≈2IDVGS−Vth,gds≈λID,(3)g_m \approx\frac{2I_D}{V_{GS}-V_{th}}, \qquad g_{ds}\approx\lambda I_D, \tag*{(3)}

where VGS−VthV_{GS}-V_{th} is the gate overdrive voltage, and λ\lambda is the channel-length-modulation parameter. These approximations rely on long-channel, strong-inversion operation in saturation and typically treat λ\lambda as constant. Their accuracy can therefore deteriorate as transistor dimensions and bias conditions change, particularly across different inversion regions.

The gain relationship in Eq. (2) illustrates the form of an analytical equation generated by the LLM. Rather than approximating gmg_m and gdsg_{ds} using Eq. (3), we evaluate the same circuit-level gain relationship using the values extracted from SPICE .OP analysis:

AvOP≈−gmOPgdsOP+1/RL.(4)A_v^{\mathrm{OP}} \approx-\frac{g_m^{\mathrm{OP}}}{g_{ds}^{\mathrm{OP}}+1/R_L}. \tag*{(4)}

The three gm/IDg_m/I_D profiles establish different transistor bias conditions and inversion regions, producing corresponding variations in the extracted gmg_m and gdsg_{ds}. These parameters reflect the transistor characteristics at each operating point without requiring the long-channel approximations in Eq. (3). Consequently, discrepancies between the analytical prediction and the reference simulation more directly indicate limitations of the circuit-level relationship, such as neglected device contributions or loading effects. This allows the subsequent repair process to focus on refining the equation rather than approximating the underlying transistor characteristics.

For each metric mm and operating-point profile kk, we evaluate the generated equation using the corresponding extracted device parameters to obtain the prediction y^m,k\hat{y}_{m,k}, while the metric-specific SPICE simulation provides the reference ym,kSPICEy_{m,k}^{\mathrm{SPICE}}. The percentage error is calculated as

em,k=100∣y^m,k−ym,kSPICE∣max⁡(∣ym,kSPICE∣,δm)%,(5)e_{m,k}=100\frac{\left|\hat{y}_{m,k}-y_{m,k}^{\mathrm{SPICE}}\right|}{\max\left(\left|y_{m,k}^{\mathrm{SPICE}}\right|,\delta_m\right)}\%, \tag*{(5)}

where δm>0\delta_m>0 is a metric-specific normalization floor that prevents excessive relative errors when the SPICE reference approaches zero. We accept an equation only if

em,k≤ϵ,∀k∈{1,2,3},(6)e_{m,k}\leq\epsilon,\qquad\forall k\in\{1,2,3\}, \tag*{(6)}

where ϵ\epsilon denotes the prescribed verification tolerance. By enforcing this criterion independently at every profile, we ensure that accuracy at one operating condition cannot compensate for inaccuracies at another. Equations that fail at any profile are forwarded to the Equation Repair Agent with their predictions, reference values, and profile-specific errors. The extracted device parameters and reference values are retained for subsequent repair iterations, avoiding repeated simulations.

Figure 7. Overview of the proposed automated analog circuit sizing framework.

Equation repair

  1. Equation repair: The Equation Repair Agent revises analytical performance equations that fail the verification criterion (Eq. (6)) for one or more operating-point profiles. We provide the LLM with the circuit netlist, functional interpretation, candidate equations, and profile-specific verification feedback, prompting it to identify neglected device contributions, loading effects, compensation-network interactions, parasitic capacitances, and bias-dependent behavior. Each proposed revision must include a physical justification for any added, removed, or modified terms. We instruct the agent to avoid arbitrary fitting coefficients while permitting numerical factors derived from circuit relationships or unit conversions, such as 0.5 for differential excitation or 2π2\pi for frequency conversion. We then evaluate each revised equation across all three profiles using previously extracted device parameters and reference values, without repeating simulations. Equations satisfying the prescribed error threshold at every profile are locked, while those failing at any profile are returned to the agent with updated errors and previous unsuccessful revisions. Repair therefore continues selectively for metrics that remain unverified.

Figure 5 illustrates the information provided to the Equation Repair Agent for the 2S-OTA. The left panel presents the diagnostic instructions and constraints governing equation revisions. The upper-right panel uses unity-gain frequency (UGF) as a representative case to compare equation predictions against SPICE across the three operating-point profiles. Equations rejected in earlier repair iterations are also provided to the agent but omitted from the figure for brevity. The lower-right panel summarizes the verification errors and status of all performance metrics, including the initial DC_gain, BW_3dB, and OCMR equations presented in § 2-C. While DC_gain passes verification and is retained, BW_3dB and OCMR require repair. The UGF errors in the detailed comparison also appear in the corresponding row of the overall summary, where a single status indicates whether the equation satisfies the verification criterion across all three profiles.

Figure 6 shows how the agent uses this feedback to progressively refine the UGF equation, presenting iterations 1, 3, and 5 from the five-iteration repair history. In iteration 1, UGF is estimated using the low-frequency gain and dominant-pole frequency at net 3, overestimating SPICE by 19.4%, 31.1%, and 39.4% across the three profiles, consistent with the initial errors in Figure 5. An intermediate two-pole formulation (not shown) substantially underestimates UGF, which the agent attributes to inconsistent frequency units (Hz versus rad/s) and excessive weighting of the compensation path. In iteration 3, the agent returns to the dominant-pole relationship ∣Av0∣f3dB|A_{v0}|f_{3\mathrm{dB}}, using the DC-gain and 3-dB-bandwidth equations verified by that iteration, and incorporates previously neglected feed-forward coupling through CcC_c, finite output-node capacitance, and second-stage gm/gdsg_m/g_{ds} loading. The resulting errors are −5.1%, 1.4%, and 3.9% at OP1, OP2, and OP3, respectively. Although OP2 and OP3 satisfy the verification criterion, OP1 remains outside the prescribed tolerance.

Based on the remaining discrepancy at OP1, the agent identifies an insufficient representation of the compensation-return path and introduces the finite drain conductance gds,M6g_{ds,M6} into the local return denominator of the CcC_c–RcR_c feed-forward path. This correction accounts for the contribution of M6 to the output-node conductance while preserving the previously verified gain and bandwidth relationships. By iteration 5, the UGF errors decrease to −3.2%, 1.3%, and 3.4% at OP1, OP2, and OP3, respectively, satisfying the verification criterion across all three profiles. The revised UGF equation is consequently locked. Upon successful verification of all requested performance metrics, we store the verified equations and their circuit context in the RAG knowledge base for future retrieval. The resulting equation set also serves as the reusable inner-loop performance evaluator in the automated sizing framework presented in the following section.

AUTOMATED SIZING FRAMEWORK

Figure 7 illustrates a framework for applying the SPICE-verified performance equations generated by NEMESIS2.0 for schematic-level circuit sizing. For a given circuit topology, the independent sizing variables, design constraints, and target specifications define the optimization problem. The sizing algorithms are based on population-based stochastic search, where a significant number of candidates must be evaluated in each iteration. To facilitate this, the equations are first taken through a Vectorization Agent (Section III-A) that transforms the verified scalar equations into numerically equivalent batch evaluators. For the circuit sizing problem (Section III-B), NEMESIS2.0 models are used to evaluate constraints, rank candidates, and guide population updates (Section III-C). The selected design is then verified against the original specifications using metric-specific SPICE testbenches [42]: if any specification is violated, the corresponding internal optimization target is adjusted based on the observed discrepancy, and sizing is repeated (Section III-D).

Vectorized model generation

A. Vectorized model generation Population-based circuit sizing requires evaluating the performance of multiple candidate designs at each optimization iteration. Although the verified scalar equations generated by NEMESIS2.0 can evaluate individual candidates, executing them separately across a population introduces the overheads of repeated function calls to the model evaluator. To improve computational efficiency, we configure a Vectorization Agent to transform these equations into batch evaluators using array-wise operations. The agent is instructed to preserve the original analytical expressions, metric definitions, numerical constants, safeguards, and operating-point-dependent conditions, modifying only their computational implementation.

For each performance metric mm, the agent transforms the verified scalar function fm,Nf_{m,\mathcal{N}} defined in Eq. (1) into a batch evaluator Fm,NF_{m,\mathcal{N}}:

y^m=Fm,N({Xd}d∈D,U),(7)\hat{y}_m = F_{m,\mathcal{N}}(\{\mathbf{X}_d\}_{d \in\mathcal{D}}, \mathbf{U}), \tag*{(7)}

where Xd\mathbf{X}_d contains the extracted device parameters of transistor dd across the candidate batch and U\mathbf{U} contains the corresponding external circuit and testbench quantities. For a batch of BB candidates, y^m\hat{y}_m contains BB predicted values of metric mm. Applying the evaluators across all requested metrics produces the performance matrix Y^\hat{\mathbf{Y}}, with candidates as rows and metrics as columns. Computational savings arise from array-wise equation evaluation rather than a reduction in the number of candidates or the operating-point analyses required to obtain their device parameters.

We verify the numerical consistency of each vectorized evaluator by comparing its predictions against those obtained from the corresponding verified scalar function using identical candidate inputs. Let y^mVec\hat{\mathbf{y}}_m^{\mathrm{Vec}} denote the vectorized predictions and y^mScalar\hat{\mathbf{y}}_m^{\mathrm{Scalar}} the predictions obtained by evaluating the scalar function independently for each candidate. We require ∣y^mVec−y^mScalar∣≤ϵabs+ϵrel∣y^mScalar∣, ∀m\left|\hat{\mathbf{y}}_m^{\mathrm{Vec}}-\hat{\mathbf{y}}_m^{\mathrm{Scalar}}\right|\leq\epsilon_{\mathrm{abs}}+\epsilon_{\mathrm{rel}}\left|\hat{\mathbf{y}}_m^{\mathrm{Scalar}}\right|,\ \forall m, where ϵabs\epsilon_{\mathrm{abs}} and ϵrel\epsilon_{\mathrm{rel}} are the prescribed absolute and relative numerical tolerances, respectively. The inequality is evaluated element-wise across the validation candidates. Any mismatch triggers correction of the vectorized implementation while the verified scalar equations remain unchanged. Evaluators satisfying this criterion on the validation inputs are accepted for circuit sizing.

Circuit sizing problem formulation

We formulate the circuit sizing problem independently of the optimization algorithm so that different optimizers can use the same design variables, constraints, and objective function. Each candidate proposed by the optimizer is represented by a vector of independent sizing variables:

θ=[θ1,θ2,…,θnθ]T,θˉ=Q(θ)∈∏j=1nθΘj,\boldsymbol{\theta}=[\theta_1,\theta_2,\ldots,\theta_{n_\theta}]^{\mathsf{T}},\qquad\bar{\boldsymbol{\theta}}=Q(\boldsymbol{\theta})\in\prod_{j=1}^{n_\theta}\Theta_j,

where nθn_\theta is the number of independent variables, which may include transistor widths, bias currents, and passive-component values. Each Θj\Theta_j defines the set of permitted values for variable θj\theta_j based on its bounds and discretization. However, the optimizer-proposed values θj\theta_j may not satisfy these requirements. Therefore, Q(⋅)Q(\cdot) applies clipping and discretization to enforce the bounds and discretization requirements, producing the valid sizing vector θˉ\bar{\boldsymbol{\theta}}.

For a batch of BB candidate designs, we instantiate each circuit using its valid sizing vector θˉ\bar{\boldsymbol{\theta}}. SPICE .OP analyses then provide the device parameters required by the performance equations. These parameters are organized into the arrays Xd\mathbf{X}_d and supplied, together with the external circuit and testbench parameters U\mathbf{U}, to the vectorized evaluators in Eq. (7). The resulting performance matrix Y^\hat{\mathbf{Y}} provides the predicted performance for all candidates.

The extracted operating-point parameters and predicted performance are then used to evaluate the corresponding constraint violations for each candidate. We first consider operating-point constraints, which impose biasing conditions on individual devices. For example, a transistor dd required to operate in saturation may be assigned the condition ∣VDS,d∣−VDSAT,d≥Vmarginsat|V_{\mathrm{DS},d}|-V_{\mathrm{DSAT},d}\geq V_{\mathrm{margin}}^{\mathrm{sat}}, where VmarginsatV_{\mathrm{margin}}^{\mathrm{sat}} is the prescribed voltage margin for saturation. For a given operating-point constraint ℓ\ell, we define the normalized violation vector vℓOP\mathbf{v}^{\mathrm{OP}}_{\ell} across candidates. For the saturation condition above, the violation for candidate ii is given by vℓ,iOP=max⁡(0,Vmarginsat−∣VDS,d,i∣+∣VDSAT,d,i∣)/VDDv^{\mathrm{OP}}_{\ell,i}=\max(0,V_{\mathrm{margin}}^{\mathrm{sat}}-|V_{\mathrm{DS},d,i}|+|V_{\mathrm{DSAT},d,i}|)/V_{\mathrm{DD}}, where VDD=1 VV_{\mathrm{DD}}=1\ \mathrm{V} is the supply voltage used for normalization. These violations are evaluated using the extracted operating-point parameters and are zero when the corresponding constraints are satisfied. Performance constraints, in contrast, are evaluated using the predicted metrics in Y^\hat{\mathbf{Y}}. For each metric mm, we define the normalized constraint-violation vector as

vm={max⁡(0,y^m−Um)sm,upper bound,max⁡(0,Lm−y^m)sm,lower bound,max⁡(0,∣y^m−Tm∣−τm)sm,target,max⁡(0,Lm−y^m,y^m−Um)sm,interval,(8)\mathbf{v}_m = \begin{cases} \dfrac{\max(0,\hat{y}_m-U_m)}{s_m}, & \text{upper bound},\\[6pt] \dfrac{\max(0,L_m-\hat{y}_m)}{s_m}, & \text{lower bound},\\[6pt] \dfrac{\max(0,|\hat{y}_m-T_m|-\tau_m)}{s_m}, & \text{target},\\[6pt] \dfrac{\max(0,L_m-\hat{y}_m,\hat{y}_m-U_m)}{s_m}, & \text{interval,} \end{cases} \tag*{(8)}

where LmL_m and UmU_m are the lower and upper specification limits, TmT_m is the target value, τm≥0\tau_m\geq0 is the designer-specified absolute tolerance defining the acceptable range [Tm−τm,Tm+τm][T_m-\tau_m,T_m+\tau_m], with its specific values provided in Section IV-B, and sm>0s_m>0 is a metric-specific normalization scale determined from the corresponding lower and upper specification bounds. The operations are applied element-wise across candidates, producing zero violation when a specification is satisfied and a positive value otherwise. Operating-point constraint violations are normalized similarly using their corresponding electrical scales.

The sizing formulation thus considers design-variable bounds enforced through Q(⋅)Q(\cdot), and operating-point and performance constraints incorporated into the objective function through their normalized violations. We combine these violations with the design objectives into an objective vector:

J=Jsoft+ρ∑m∈Cperfhardwm(vm⊙vm)+ρ∑ℓ∈COPhardwℓ(vℓOP⊙vℓOP),(9)\mathbf{J}=\mathbf{J}_{\mathrm{soft}}+\rho\sum_{m\in\mathcal{C}_{\mathrm{perf}}^{\mathrm{hard}}}w_m(\mathbf{v}_m\odot\mathbf{v}_m)+\rho\sum_{\ell\in\mathcal{C}_{\mathrm{OP}}^{\mathrm{hard}}}w_\ell(\mathbf{v}^{\mathrm{OP}}_\ell\odot\mathbf{v}^{\mathrm{OP}}_\ell), \tag*{(9)}

where Jsoft\mathbf{J}_{\mathrm{soft}} contains the configured design objectives, expressed as minimization terms. The sets Cperfhard\mathcal{C}_{\mathrm{perf}}^{\mathrm{hard}} and COPhard\mathcal{C}_{\mathrm{OP}}^{\mathrm{hard}} contain the mandatory performance and operating-point constraints, respectively. The coefficient ρ>0\rho>0 controls the penalty contribution, while wmw_m and wℓw_\ell are the corresponding nonnegative weights.

Population-based optimizers

We employ population-based optimizers to determine the sizing parameters of a given circuit topology. In iteration tt, each optimizer maintains a population of NpN_p candidate designs:

P(t)={θ1(t),θ2(t),…,θNp(t)},(10)\mathcal{P}^{(t)} = \left\{\theta_1^{(t)}, \theta_2^{(t)}, \ldots, \theta_{N_p}^{(t)}\right\}, \tag*{(10)}

where θi(t)\theta_i^{(t)} contains the independent sizing variables of candidate ii. We process each new or modified candidate using the parameter mapping and constraint-evaluation procedure described in Section III-B, supplying the extracted device parameters to the vectorized performance equations. Unchanged candidates reuse their cached operating-point parameters, performance predictions, and objective values. We represent the resulting candidate evaluation as

(Ji,χi,y^i,vi)=E(θi),(11)(J_i, \chi_i, \hat{\mathbf{y}}_i, \mathbf{v}_i) = \mathcal{E}(\theta_i), \tag*{(11)}

where E(⋅)\mathcal{E}(\cdot) denotes the candidate-evaluation process, JiJ_i is the objective value, χi\chi_i is the hard-feasibility indicator, y^i\hat{\mathbf{y}}_i contains the predicted performance metrics, and vi\mathbf{v}_i contains the normalized constraint violations.

Each optimizer uses these evaluation results to update its candidate population according to its optimization strategy. The evaluation–update cycle continues until a prescribed stopping condition is reached, such as obtaining a satisfactory feasible design or exhausting the evaluation budget. We retain the best feasible candidate encountered, based on its objective value. If no candidate satisfies all mandatory constraints, the optimization run is considered unsuccessful.

We consider three population-based optimizers: GA [36], PSO [37], and DE [38]. GA generates new sizing configurations through selection, crossover, and mutation; PSO updates candidate positions based on individual and swarm-best solutions; and DE constructs trial sizing vectors using scaled differences between population members. All three algorithms use the same circuit-evaluation procedure, sizing constraints, and objective function, enabling a consistent comparison of their performance under identical evaluation budgets.

Final SPICE-based verification

Once the optimizer identifies a feasible design using the analytical performance model, we instantiate the optimized circuit and evaluate its performance using metric-specific SPICE testbenches [42]. Unlike the earlier equation-verification stage, which assesses model accuracy across representative operating-point profiles, this final verification determines whether the optimized circuit satisfies the original application specifications. We compare the simulated performance with both the specifications and the model predictions to identify any discrepancies at the optimized operating point.

If any specification is violated, we tighten the corresponding internal optimization target based on the observed discrepancy and repeat the sizing process using the same verified vectorized equations. The resulting candidate is then reevaluated in SPICE. This verification–retuning cycle continues until all specifications are satisfied or the allotted optimization budget is exhausted. Only designs that meet the original specifications in the final SPICE simulations are considered successful. The SPICE-verified sizing results are presented in Section IV.

Experiments and Results

Experimental setup

We evaluate NEMESIS2.0 and its automated sizing framework using seven circuit testcases comprising five OTAs and two COMPs. Table I summarizes the testcases and their modeled performance metrics. All circuits are implemented in a commercial 65 nm CMOS technology with a 1 V supply and simulated using Synopsys HSPICE. To examine model adaptability to external loading conditions, we also evaluate the generated equations across output load capacitances (CLC_L) ranging from 100 fF to 1 pF.

Circuit class / TopologiesPerformance metrics
OTAs
OTA-1: 5-transistor OTA
OTA-2: Current-mirror OTA
OTA-3: 2-stage OTA
OTA-4: Low-voltage cascode OTA
OTA-5: Telescopic OTA
AC: DC gain (ADCA_{\mathrm{DC}}), Unity-gain frequency (UGF), 3-dB bandwidth (BW3dB\mathrm{BW}_{3\mathrm{dB}}), Phase margin (PM), Common-mode Rejection ratio (CMRR), Power-supply rejection ratio (PSRR+/-)
DC sweep: Input and output common-mode range (ICMRmin⁡/max⁡\mathrm{ICMR}_{\min/\max} and OCMRmin⁡/max⁡\mathrm{OCMR}_{\min/\max})
Transient: Slew rate (SR+/-)
Power: DC power consumption (PDCP_{\mathrm{DC}})
Area: Active-device area (AdevA_{\mathrm{dev}})
Comparator
COMP-1: StrongARM comparator
COMP-2: Double-tail comparator
Transient: Evaluation-delay (tevalt_{\mathrm{eval}}), Energy per decision (EdecE_{\mathrm{dec}}), Precharge delay (tpret_{\mathrm{pre}})
Transient sweep: ICMRmin⁡/max⁡\mathrm{ICMR}_{\min/\max}
Area: Active-device area (AdevA_{\mathrm{dev}})

Table I. Circuit testcases and modeled performance metrics.

All LLM-based agents use OpenAI GPT-5.5 [43] with high reasoning effort. We repeat equation generation and repair over five independent runs per circuit and report averaged results unless otherwise specified. The equation-verification tolerance in Eq. (6) is set to ϵ=5%\epsilon= 5\% for all modeled metrics. Our experiments assess the accuracy and convergence of the generated equations and their effectiveness in automated sizing under different performance targets. For comparisons across GA, PSO, and DE, we maintain identical sizing variables, bounds, constraints, initialization, and candidate-evaluation budgets. The open-source repository is available at https://github.com/UMN-EDA/NEMESIS.

Performance model accuracy and convergence

We evaluate the quality of the generated performance equations across the seven circuit testcases from three complementary perspectives. First, we assess model generalization by comparing equation predictions against SPICE references at previously unseen sizing configurations and quantifying the resulting prediction errors. Second, we examine these errors relative to metric-specific tolerances, allowing accuracy to be compared across performance metrics with different units and numerical scales. Third, we evaluate the modeling effort by examining the number of iterations required to obtain verified equations and their final verification errors.

Model generalization and error statistics

1) Model generalization and error statistics: To evaluate model generalization, we examine the predictive accuracy of the verified equations beyond the three representative operating-point profiles used during model construction. For each testcase, we generate an additional 1000 fully sized circuit candidates using Latin hypercube sampling (LHS) [44], covering a wider range of transistor dimensions than those used during equation construction. These candidates are excluded from equation generation and repair, allowing us to assess the fixed equations at previously unseen operating points. We compare the predicted performance metrics against their corresponding SPICE references for each candidate.

For performance metric mm and LHS candidate ii, the signed model–SPICE residual is defined as

Δm,i=ym,iModel−ym,iSPICE\Delta_{m,i}=y^{\mathrm{Model}}_{m,i}-y^{\mathrm{SPICE}}_{m,i}

where ym,iModely^{\mathrm{Model}}_{m,i} and ym,iSPICEy^{\mathrm{SPICE}}_{m,i} denote the model prediction and SPICE reference, respectively. Positive residuals indicate overestimation, while negative residuals indicate underestimation. The blue-shaded columns in Table II and Table III summarize the prediction errors for the OTA and comparator testcases, respectively, using the following statistics:

MetricOTA-1 μ\muOTA-1 σ\sigmaOTA-1 MAEOTA-1 kfk_fOTA-1 ϵf\epsilon_fOTA-2 μ\muOTA-2 σ\sigmaOTA-2 MAEOTA-2 kfk_fOTA-2 ϵf\epsilon_fOTA-3 μ\muOTA-3 σ\sigmaOTA-3 MAEOTA-3 kfk_fOTA-3 ϵf\epsilon_fOTA-4 μ\muOTA-4 σ\sigmaOTA-4 MAEOTA-4 kfk_fOTA-4 ϵf\epsilon_fOTA-5 μ\muOTA-5 σ\sigmaOTA-5 MAEOTA-5 kfk_fOTA-5 ϵf\epsilon_fTolerance
ADCA_\mathrm{DC} (dB)-0.230.080.2312.070.600.130.6014.31-0.290.130.2910.100.871.070.8731.410.302.301.5120.87Abs., 2 dB
UGF (MHz)0.330.120.3311.490.160.260.2183.232.014.753.1053.420.231.070.2533.940.071.230.8124.37Rel., 5%
BW3dB_{3\mathrm{dB}} (MHz)0.000.010.0131.750.020.010.0281.760.160.810.2164.970.231.070.2530.900.994.031.0543.26Rel., 5%
PM (∘^\circ)-0.050.080.0710.190.070.260.2120.95-4.214.684.6064.12-5.444.585.4414.15-3.259.875.2722.27Abs., 5∘5^\circ
CMRR (dB)1.090.561.0954.900.681.301.1722.74-2.292.642.7362.911.094.022.8442.23-2.454.523.9833.98Abs., 5 dB
PSRR+ (dB)-0.120.060.1221.280.401.080.9124.84-1.151.461.5944.941.362.702.2744.03-2.022.162.8724.26Abs., 2 dB
PSRR− (dB)-0.760.981.0322.80-0.110.270.2122.02-2.045.894.0634.412.475.305.7544.36-2.056.496.3134.75Abs., 2 dB
ICMRmin_\mathrm{min} (mV)20.7435.4232.3543.361.8633.1326.2844.7812.6817.2416.3654.4633.7030.3138.0234.8625.3435.6534.1634.62Abs., 50 mV
ICMRmax_\mathrm{max} (mV)38.9023.4839.2014.682.7946.6834.7444.89-31.5045.7237.3762.90-27.9161.2246.7734.241.9537.0531.5824.78Abs., 50 mV
OCMRmin_\mathrm{min} (mV)15.6021.1520.1354.08-0.723.762.4544.64-25.77143.4033.27134.9120.4718.9022.1843.4817.4022.7620.2222.36Abs., 50 mV
OCMRmax_\mathrm{max} (mV)-29.0924.3633.4723.036.3124.9421.1722.91-44.1053.6646.9454.6722.1437.4633.0112.83-25.2636.5634.8824.48Abs., 50 mV
SR+ (V/μ\mus)1.423.482.6654.960.932.511.6862.375.358.957.05114.770.800.730.8142.701.372.311.6633.50Rel., 10%
SR− (V/μ\mus)2.083.242.7064.560.842.541.7863.935.294.835.84134.922.332.142.3644.900.070.900.7031.80Rel., 10%

Table II. OTA model results: accuracy (μ\mu, σ\sigma, MAE), error tolerances, and convergence (kfk_f, ϵf\epsilon_f).

MetricCOMP-1 μ\muCOMP-1 σ\sigmaCOMP-1 MAECOMP-1 kfk_fCOMP-1 ϵf\epsilon_f (%)COMP-2 μ\muCOMP-2 σ\sigmaCOMP-2 MAECOMP-2 kfk_fCOMP-2 ϵf\epsilon_f (%)Tolerance
tevalt_\mathrm{eval} (ps)-10.6717.8016.10124.8511.8013.0015.0063.72Rel., 10%
EdecE_\mathrm{dec} (fJ)14.226.4715.57313.2342.0427.7842.1491.83Rel., 10%
ICMRmin_\mathrm{min} (mV)5.066.155.0924.31-4.0826.3816.1324.77Abs., 50 mV
ICMRmax_\mathrm{max} (mV)-5.1220.9310.87301.113.320.473.3220.39Abs., 50 mV
tpret_\mathrm{pre} (ps)-10.6331.5124.6974.0210.7936.0011.3094.62Rel., 10%

Table III. Comparator model results: accuracy (μ\mu, σ\sigma, MAE), error tolerances, and convergence (kfk_f, ϵf\epsilon_f).

μm=1Nm∑i=1NmΔm,i,σm=1Nm−1∑i=1Nm(Δm,i−μm)2,\mu_m=\frac{1}{N_m}\sum_{i=1}^{N_m}\Delta_{m,i},\qquad\sigma_m=\sqrt{\frac{1}{N_m-1}\sum_{i=1}^{N_m}(\Delta_{m,i}-\mu_m)^2},
MAEm=1Nm∑i=1Nm∣Δm,i∣.\mathrm{MAE}_m=\frac{1}{N_m}\sum_{i=1}^{N_m}\lvert\Delta_{m,i}\rvert.

Here, NmN_m denotes the number of evaluated LHS candidates for metric mm. The mean residual (μm\mu_m) indicates systematic overestimation or underestimation, while the standard deviation (σm\sigma_m) measures the spread of prediction errors. The mean absolute error (MAE) quantifies the average error magnitude regardless of its sign. All three statistics are expressed in the units of the corresponding performance metric.

The results illustrate differences in prediction accuracy across metrics and circuit topologies. The signed mean residual complements MAE by identifying systematic overestimation or underestimation, which cannot be determined from absolute errors alone. For example, the DC-gain model of OTA-3 exhibits μ=−0.29\mu=-0.29 dB and MAE =0.29=0.29 dB, indicating predominantly negative residuals and hence underestimation. In contrast, the SR+ model of OTA-2 has μ=0.93\mu=0.93 V/μ\mus, σ=2.51\sigma=2.51 V/μ\mus, and MAE =1.68=1.68 V/μ\mus. The difference between the signed mean and MAE reflects partial cancellation of positive and negative residuals, while the positive mean indicates an overall tendency toward overestimation.

Normalized model vs. SPICE error analysis

The preceding error statistics quantify prediction accuracy in the units of individual performance metrics, making direct comparisons across metrics and circuit topologies difficult. To establish a common scale, we

normalize each signed model–SPICE residual by its metric-specific error tolerance:

rm,i=Δm,iηm,i,ηm,i={τm,absolute tolerance,ρm∣ym,iSPICE∣,relative tolerance.r_{m,i}=\frac{\Delta_{m,i}}{\eta_{m,i}},\qquad\eta_{m,i}= \begin{cases} \tau_m, & \text{absolute tolerance},\\ \rho_m\lvert y^{\mathrm{SPICE}}_{m,i}\rvert, & \text{relative tolerance}. \end{cases}

where ηm,i\eta_{m,i} is the metric-specific error-normalization scale, τm\tau_m is the absolute tolerance, and ρm\rho_m is the relative tolerance factor. The green-shaded columns in Table II and Table III specify these tolerances for each metric. Absolute tolerances are used for ADCA_\mathrm{DC}, PM, CMRR, PSRR+/−+/-, ICMRmin/max_\mathrm{min/max}, and OCMRmin/max_\mathrm{min/max}, since their prediction errors are more meaningfully interpreted as differences in dB, degrees, or voltage. Relative tolerances are used for UGF, BW3dB_{3\mathrm{dB}}, SR+/−+/-, EdecE_\mathrm{dec}, tevalt_\mathrm{eval}, and tpret_\mathrm{pre}, for which fractional deviations from the reference values provide a meaningful measure of error. These tolerances represent acceptable prediction errors for circuit sizing and are distinct from the 5% equation-verification threshold used during model construction. Accordingly, rm,i=0r_{m,i}=0 indicates exact agreement, ∣rm,i∣≤1\lvert r_{m,i}\rvert\le1 indicates an error within tolerance, and ∣rm,i∣>1\lvert r_{m,i}\rvert> 1 indicates a tolerance exceedance.

Figure 8 presents the normalized residual distributions for the OTA metrics. The solid line at rm,i=0r_{m,i}=0 indicates exact model–SPICE agreement, the dashed lines at ±1\pm1 mark the tolerance limits, and the thick red lines within the boxes represent mean residuals. Most distributions are centered near zero, although SR+/−^{+/-} and PSRR+/−^{+/-} exhibit comparatively larger variations. Here, AdevA_{\mathrm{dev}} and PDCP_{\mathrm{DC}} have zero residual across the evaluated candidates and are represented by diamond markers at rm,i=0r_{m,i}=0 and Adev=∑d∈DWdLdA_{\mathrm{dev}}=\sum_{d\in\mathcal{D}}W_dL_d provides an area estimate at the schematic stage.

Distribution of the OTA model–SPICE residuals

Figure 8. Distribution of the OTA model–SPICE residuals relative to their metric-specific error tolerances.

Model convergence

The orange-shaded columns in Table II and Table III summarize equation-repair convergence. For each metric, kfk_f denotes the repair iteration attaining the lowest recorded maximum verification error across the three operating-point profiles, and εf\varepsilon_f is the corresponding error. All OTA and COMP equations satisfy εf<5%\varepsilon_f<5\%, confirming successful verification under the prescribed criterion. However, the refinement effort varies across circuits and metrics. For OTA-3, OCMRmin⁡_{\min} and SR−^- reach kf=13k_f=13, whereas all COMP-2 metrics converge within nine iterations. COMP-1 requires more extensive refinement for selected metrics, reaching kf=31k_f=31 for EdecE_{\mathrm{dec}} and kf=30k_f=30 for ICMRmax⁡_{\max}. These results show that convergence behavior depends on the circuit topology and performance metric, even under a common verification criterion.

Model-driven sizing evaluation

We evaluate the generated performance equations for automated sizing across seven testcases through three experiments. First, we perform application-specific sizing to examine performance tradeoffs under different optimization objectives. Second, we assess single-sizing-attempt feasibility across 100 specification sets per testcase and evaluate the effectiveness of target tightening in recovering from unsuccessful attempts. Third, we compare GA, PSO, and DE using the same equations to assess optimizer portability and quantify runtime savings relative to SPICE-in-the-loop sizing. We independently verify the resulting designs against their original specifications.

Application-specific sizing

We employ GA to size the seven testcases in Table I under application-specific performance objectives. For each OTA, we consider four sizing configurations: high speed, low power, balanced, and high gain. All four configurations use the same verified equations. The COMPs are sized under timing and input common-mode requirements, with energy and area minimization as additional objectives. Table IV and Table V present the target specifications alongside the final SPICE-verified results. Numerical specification bounds are enforced as hard constraints, while metrics marked “min.” or “max.” are treated as soft optimization objectives, following Section III-C. We additionally enforce transistor turn-on and saturation constraints for NMOS and PMOS devices: ∣VGS∣>∣VTH∣\lvert V_{\mathrm{GS}}\rvert>\lvert V_{\mathrm{TH}}\rvert, ∣VDS∣>∣VGS∣−∣VTH∣\lvert V_{\mathrm{DS}}\rvert>\lvert V_{\mathrm{GS}}\rvert-\lvert V_{\mathrm{TH}}\rvert. These device-level constraints are omitted from the tables for brevity.

MetricHigh speed TargetHigh speed SPICELow power TargetLow power SPICEBalanced TargetBalanced SPICEHigh gain TargetHigh gain SPICE
ADCA_{\mathrm{DC}} (dB)≥20\ge2021.73≥20\ge2021.05≥22\ge2222.68≥27\ge2728.44
UGF (MHz)≥600\ge600683.66≥200\ge2003.21≥80\ge8088.54≥100\ge1006.70
BW3dB_{3\mathrm{dB}} (MHz)≥50\ge5055.13max.0.29≥70\ge706.05max.0.28
PM (∘^\circ)≥60\ge6088.76≥60\ge6083.14≥70\ge7085.05≥70\ge7080.25
CMRR (dB)≥45\ge4548.79max.48.43≥45\ge4548.24≥55\ge5560.56
PSRR+^+ (dB)max.20.91max.21.89max.22.46≥25\ge2528.34
PSRR−^- (dB)max.46.20max.52.28max.52.71≥25\ge2560.56
ICMRmin⁡_{\min} (V)≤0.4\le0.40.39≤0.4\le0.40.39≤0.4\le0.40.40≤0.4\le0.40.35
ICMRmax⁡_{\max} (V)≥0.8\ge0.80.87≥0.8\ge0.80.88≥0.8\ge0.80.85≥0.8\ge0.80.85
OCMRmin⁡_{\min} (V)≤0.3\le0.30.22≤0.2\le0.20.18≤0.2\le0.20.17≤0.2\le0.20.19
OCMRmax⁡_{\max} (V)≥0.8\ge0.80.82≥0.8\ge0.80.85≥0.8\ge0.80.85≥0.8\ge0.80.86
SR+^+ (V/μ\mus)≥300\ge300355.81max.12.57≥40\ge4042.79max.3.85
SR−^- (V/μ\mus)≥300\ge300379.12max.11.42≥40\ge4041.01max.5.28
PDCP_{\mathrm{DC}} (μ\muW)min.523.55≤10\le109.73min.42.97min.29.26
AdevA_{\mathrm{dev}} (μ\mum2^2)min.34.56min.24.08min.49.76min.232.08
tft_f (sec)13.5819.5218.0415.26
OTA-2 (CM-OTA)
ADCA_{\mathrm{DC}} (dB)≥20\ge2020.56≥20\ge2023.54≥22\ge2222.60≥30\ge3030.54
UGF (MHz)≥490\ge490511.49max.26.03≥80\ge80138.00max.1.07
BW3dB_{3\mathrm{dB}} (MHz)≥50\ge5051.75max.1.73≥15\ge1510.23max.0.031
PM (∘^\circ)≥60\ge6093.32≥60\ge6094.05≥70\ge7094.69≥70\ge7093.81
CMRR (dB)≥45\ge4551.62max.53.54≥45\ge4550.31≥55\ge5558.04
PSRR+^+ (dB)max.29.12max.58.01max.58.08≥50\ge5055.76
PSRR−^- (dB)max.33.24max.25.50max.24.89≥35\ge3539.34
ICMRmin⁡_{\min} (V)≤0.5\le0.50.43≤0.26\le0.260.26≤0.27\le0.270.27≤0.3\le0.30.28
ICMRmax⁡_{\max} (V)≥0.8\ge0.80.84≥0.7\ge0.70.75≥0.8\ge0.80.87≥0.8\ge0.80.84
OCMRmin⁡_{\min} (V)≤0.2\le0.20.17≤0.2\le0.20.18≤0.2\le0.20.19≤0.2\le0.20.24
OCMRmax⁡_{\max} (V)≥0.8\ge0.80.88≥0.7\ge0.70.88≥0.8\ge0.80.89≥0.8\ge0.80.86
SR+^+ (V/μ\mus)≥800\ge800874.27max.11.49≥20\ge2041.08max.86.96
SR−^- (V/μ\mus)≥800\ge800809.04max.12.45≥30\ge3055.85max.86.97
PDCP_{\mathrm{DC}} (μ\muW)min.1543.79≤30\le3028.42min.94.89min.364.80
AdevA_{\mathrm{dev}} (μ\mum2^2)min.15.24min.55.00min.54.50min.290.24
tft_f (sec)22.7321.4414.0927.29
OTA-3 (2S-OTA)
ADCA_{\mathrm{DC}} (dB)≥25\ge2527.04≥25\ge2529.07≥40\ge4044.92≥60\ge6062.72
UGF (MHz)≥500\ge500582.97max.28.40≥80\ge80108.58max.42.90
BW3dB_{3\mathrm{dB}} (MHz)≥20\ge2025.03max.1.39≥1\ge11.08max.0.08
PM (∘^\circ)≥60\ge6089.19≥60\ge6070.60≥70\ge7082.95≥60\ge6065.87
CMRR (dB)≥45\ge4549.62≥45\ge4555.27≥45\ge4560.09≥55\ge5557.04
PSRR+^+ (dB)max.21.26max.34.75max.37.38≥40\ge4043.48
PSRR−^- (dB)max.54.68max.34.80max.50.14≥50\ge5059.05
ICMRmin⁡_{\min} (V)≤0.4\le0.40.38≤0.4\le0.40.33≤0.3\le0.30.28≤0.3\le0.30.27
ICMRmax⁡_{\max} (V)≥0.7\ge0.70.74≥0.8\ge0.80.88≥0.8\ge0.80.88≥0.8\ge0.80.92
OCMRmin⁡_{\min} (V)≤0.2\le0.20.12≤0.2\le0.20.17≤0.2\le0.20.19≤0.3\le0.30.28
OCMRmax⁡_{\max} (V)≥0.7\ge0.70.79≥0.8\ge0.80.82≥0.8\ge0.80.83≥0.8\ge0.80.89
SR+^+ (V/μ\mus)≥400\ge400488.16≥0.8\ge0.87.03≥30\ge3036.13max.4.63
SR−^- (V/μ\mus)≥400\ge400475.76max.4.00≥30\ge3034.56max.2.94
PDCP_{\mathrm{DC}} (μ\muW)min.1163.66≤20\le2018.55min.86.92min.113.27
AdevA_{\mathrm{dev}} (μ\mum2^2)min.10.18min.14.84min.31.16min.298.00
tft_f (sec)13.8120.6818.0717.22

Table IV.

MetricStrongARM TargetStrongARM SPICEDouble-Tail TargetDouble-Tail SPICE
tevalt_{\mathrm{eval}} (ps)≤50\leq5034.81≤50\leq5044.36
tpret_{\mathrm{pre}} (ps)≤100\leq10076.45≤40\leq4038.57
EdecE_{\mathrm{dec}} (fJ)min.193.29min.52.73
ICMRmin⁡_{\min} (V)≤0.4\leq0.40.34≤0.4\leq0.40.36
ICMRmax⁡_{\max} (V)≥0.7\geq0.70.75≥0.7\geq0.70.79
AdevA_{\mathrm{dev}} (µm2^{2})min.4.66min.25.35
tft_f (sec)21.7622.57

Table V. GA-based COMP sizing.

The OTA results in Table IV illustrate the tradeoffs associated with the four sizing configurations. High-speed sizing achieves higher UGF, BW3dB_{3\mathrm{dB}}, and SR+/−^{+/-} at increased power consumption, whereas low-power sizing reduces PDCP_{\mathrm{DC}} with lower speed. Balanced sizing addresses simultaneous gain, stability, signal-range, and speed requirements, while high-gain sizing increases ADCA_{\mathrm{DC}}, generally with reduced bandwidth and increased AdevA_{\mathrm{dev}}. For example, OTA-3 achieves a UGF of 583 MHz at 1164 µW in the high-speed configuration, compared with 28.4 MHz at 18.6 µW in the low-power configuration. Its high-gain configuration reaches 62.7 dB, with an active-device area of 298 µm2^{2}. For the comparators, Table V shows that both designs satisfy their tevalt_{\mathrm{eval}}, tpret_{\mathrm{pre}}, and ICMR specifications, with different outcomes for EdecE_{\mathrm{dec}} and AdevA_{\mathrm{dev}}. Across all 22 application-specific sizing tasks, the selected designs satisfy every specification in the first SPICE-based verification. No additional sizing passes or specification tightening are required. Each optimization completes within 34 s, followed by 7 s for final verification.

Specification-stress and recovery evaluation

To evaluate specification satisfaction beyond the application-specific sizing tasks, we consider 100 distinct specification sets per testcase, totaling 700 sizing tasks. Table VI reports the number of attempts required to satisfy the specifications under SPICE-based verification. Of the 700 tasks, 583 (83.3%) achieve feasibility on the first attempt, while the remaining tasks succeed within five attempts. The last two columns report the average cumulative runtime (tˉf\bar{t}_{f}) and candidate evaluations (NeN_{e}) across all attempts until feasibility is achieved. Runtime includes optimization and verification for each attempt.

Circuit12345tˉf\bar{t}_{f} (s)NeN_{e}
OTA-1841231027.3513553
OTA-2811161135.5913991
OTA-385931229.5312882
OTA-487822137.4812177
OTA-578974251.0715254
COMP-186751134.4210541
COMP-282692138.2811991

Table VI. Sizing efficiency across 100 specification sets.

For tasks failing initial verification, we tighten the model-side targets of the failed metrics and repeat sizing without modifying the equations. Table VII illustrates this process for OTA-3. The green entries in Attempt 1 identify the original specifications, red values mark specification violations, and bold targets indicate subsequent tightening. The first attempt fails UGF, BW3dB_{3\mathrm{dB}}, ICMRmin⁡_{\min}, and SR+/−^{+/-}. After tightening, Attempt 2 satisfies UGF and SR+^{+}, while the other three metrics remain infeasible. Further tightening produces a design satisfying all original specifications in Attempt 3, requiring a cumulative runtime of 125.7 s and 30,081 candidate evaluations. Feasibility is assessed against the original specifications, not the tightened model-side targets. For example, the final BW3dB_{3\mathrm{dB}} of 637.6 MHz satisfies the original 600 MHz requirement despite falling below the tightened 720 MHz target. Thus, target tightening enables recovery from initial verification failures without regenerating the performance equations.

MetricAttempt 1 TargetAttempt 1 SPICEAttempt 2 TargetAttempt 2 SPICEAttempt 3 TargetAttempt 3 SPICERuntime
ADCA_{\mathrm{DC}} (dB)≥25.0\geq25.034.11≥25.0\geq25.030.49≥25\geq2529.71Attempt 1: 19.8 SPICE: 6.9
UGF (GHz)≥14\geq1412.6≥14.2\geq14.214.2≥14.2\geq14.214.5Attempt 2: 41.3 SPICE: 6.9
BW3dB_{3\mathrm{dB}} (MHz)≥600\geq600219.7≥700\geq700383.6≥720\geq720637.6Attempt 3: 43.9 SPICE: 6.9
PM (°)≥60.0\geq60.089.57≥60.0\geq60.089.53≥60.0\geq60.089.50Total 125.7
CMRR (dB)≥55\geq5551.34≥50.0\geq50.057.88≥50.0\geq50.052.06
ICMRmin⁡_{\min} (V)≤0.40\leq0.400.50≤0.35\leq0.350.45≤0.30\leq0.300.39
ICMRmax⁡_{\max} (V)≥0.70\geq0.700.71≥0.70\geq0.700.72≥0.70\geq0.700.75
OCMRmin⁡_{\min} (V)≤0.10\leq0.100.03≤0.10\leq0.100.03≤0.10\leq0.100.03
OCMRmax⁡_{\max} (V)≥0.60\geq0.600.62≥0.60\geq0.600.66≥0.60\geq0.600.66
SR+^{+} (V/µs)≥900\geq900870.8≥950\geq950990.9≥950\geq950989.7
SR−^{-} (V/µs)≥900\geq900785.3≥950\geq950847.0≥1000\geq1000904.5
PDCP_{\mathrm{DC}} (µW)min.2450min.2060min.2491
Area (µm2^{2})min.36.98min.40.94min.48.50

Table VII. Representative spec-tightening recovery for OTA-3.

Optimizer portability

To evaluate optimizer portability, we apply GA, PSO, and DE to all seven testcases using the same verified performance equations. The optimizers share identical design variables, bounds, specifications, operating-region constraints, initial populations, and evaluation budgets. Each executes 35 generations with 385 candidates per generation, totaling Ne=13,475N_{e}=13{,}475 evaluations. Table VIII presents the resulting designs for a representative high-gain OTA-3 sizing task. GA achieves the highest BW3dB_{3\mathrm{dB}} (545.9 kHz), PSO achieves the highest UGF (47.14 MHz), while DE yields the lowest PDCP_{\mathrm{DC}} (43.90 µW) and AdevA_{\mathrm{dev}} (174.58 µm2^{2}). These results show that the three optimizers produce different sizing outcomes using the same equation-based evaluator, without equation regeneration or optimizer-specific modifications.

Opt.ADCA_{\mathrm{DC}} (dB)BW (kHz)UGF (MHz)CMRR (dB)PSRR+^{+}/PSRR−^{-} (dB)PDCP_{\mathrm{DC}} (µW)Area (µm2^{2})tft_f (sec)
GA60.06545.9043.2775.9049.96/68.0863.46281.8030.94
PSO60.05198.7047.1485.7746.23/68.1967.34276.1524.41
DE60.10165.0339.5247.4041.71/66.2243.90174.5816.50
Ne=13,475N_e=13{,}475 per optimizer (385 candidates × 35 generations)

Table VIII. Optimizer portability with the equation-based performance model for sizing OTA-3.

Runtime analysis

We compare NEMESIS2.0-based sizing with SPICE-in-the-loop sizing across all seven testcases. For each specification set and optimizer, both approaches use identical design variables, initialization, optimization settings, evaluation budgets, and computing resources. All runs complete the prescribed 13,475 candidate evaluations, regardless of when feasibility is reached. SPICE-in-the-loop sizing executes metric-specific simulations for every candidate, with candidate evaluations parallelized across multiple CPU cores, whereas model-driven sizing evaluates the vectorized equations using extracted device parameters, followed by SPICE verification of the selected design. Table IX reports the mean runtimes and speedups. For a single sizing attempt, the SPICE-in-the-loop baseline requires 112–115 min, compared with 28.74–34.24 s for NEMESIS2.0, yielding speedups of 196.26–237.99×.

ExperimentEvaluation effortGA RuntimeGA SpeedupPSO RuntimePSO SpeedupDE RuntimeDE Speedup
SPICE-in-the-loop sizing13,475 candidates112 minn/a115 minn/a114 minn/a
NEMESIS2.0-based sizing (One attempt)13,475 candidates + 1 SPICE verification34.24 sec196.26×30.69 sec224.82×28.74 sec237.99×
NEMESIS2.0-based sizing (Five attempts)67,375 candidates + 5 SPICE verifications173.65 sec38.69×143.78 sec47.98×137.11 sec49.88×

Table IX. Runtime speedup against SPICE-in-the-loop sizing.

The table also includes a five-attempt recovery scenario, which achieves speedups of 38.69–49.88× relative to the single-attempt SPICE-in-the-loop baseline. Model-generation runtime is excluded from these comparisons because the equations are generated once per circuit and reused across subsequent sizing tasks. Each equation-repair iteration requires 3–5 min, primarily due to LLM latency and server load.

Importance of multimodality

We evaluate how schematic information affects the identification of circuit structures and device roles by the Topology Analysis Agent. For each of the seven testcases, we perform 10 independent evaluations per input configuration, using either the netlist alone or the netlist with the schematic, while keeping all other inference settings identical. Table X compares

CircuitNetlist only TopologyNetlist only PrimitivesNetlist only RolesNetlist only NetsNetlist + schematic TopologyNetlist + schematic PrimitivesNetlist + schematic RolesNetlist + schematic Nets
OTA-130.080.076.393.8100.0100.0100.0100.0
OTA-250.086.786.086.7100.0100.0100.0100.0
OTA-360.081.470.092.2100.0100.0100.0100.0
OTA-470.082.078.388.9100.0100.0100.0100.0
OTA-520.041.768.147.5100.0100.0100.0100.0
COMP-180.087.177.086.3100.0100.0100.0100.0
COMP-260.077.170.081.0100.0100.0100.0100.0

Table X. Ablation study on circuit structure identification accuracy (%) using netlist-only vs. netlist+schematic input.

the identification accuracy against ground-truth annotations across four categories: topology, primitives, device roles, and nets. Topology accuracy is the percentage of evaluations correctly identifying the circuit topology, while the remaining scores represent the average percentage of correctly identified primitives, device roles, and nets. The net-identification score includes circuit ports, signal paths, and connectivity.

With netlist-only input, topology identification accuracy ranges from 20% to 80%, with additional errors in primitive recognition, device-role assignment, and net interpretation. For example, OTA-3 achieves 60% topology accuracy, 81.4% primitive accuracy, 70.0% role accuracy, and 92.2% net accuracy. Incorporating the schematic increases all four scores to 100% for OTA-3, with the same result observed across the remaining six testcases. Thus, although the netlist contains the required connectivity information, its textual representation alone does not consistently yield the correct circuit interpretation. The schematic provides additional structural context that eliminates the observed identification errors across the evaluated runs, establishing the benefit of multimodal inputs in circuit interpretation for performance equation generation.

Conclusion

This work presented NEMESIS2.0, an open-source agentic framework that automatically generates interpretable performance equations for analog circuits and applies them to automated sizing. The framework combines multimodal circuit interpretation, graph-based knowledge retrieval, and SPICE-guided equation repair to construct topology-specific analytical models without manual derivation. These equations are integrated with population-based optimization algorithms to size circuits under different performance objectives. Evaluated across five OTA and two COMP testcases in a commercial 65 nm CMOS technology, the generated equations achieve errors within 5% of SPICE. All 22 application-specific sizing tasks satisfy hard specifications on the first SPICE verification. Across 700 specification-stress tasks, 83.3% achieve feasibility in the first attempt, with the remainder succeeding within five attempts through target tightening. For single-attempt sizing, the framework achieves 196–238× speedup over SPICE-in-the-loop sizing under identical evaluation budgets. These results show that automatically generated analytical equations enable efficient, application-specific analog sizing across circuit topologies and optimization algorithms while retaining independent SPICE verification.

Acknowledgment

The framework incorporates LLMs, specifically OpenAI GPT-5.5, as a core modeling component, as detailed in § 2. Beyond their role in the methodology, the authors used OpenAI ChatGPT to assist with the initial development of the software environment and with editorial revisions to improve the clarity and readability of the manuscript. All AI-assisted contributions were reviewed and verified by the authors.

References

Biographies

Subhadip Ghosh

Subhadip Ghosh received the B.Tech. degree in electronics and telecommunication engineering from the Kalinga Institute of Industrial Technology, Bhubaneswar, India, in 2021, and the M.S. degree in electrical engineering from the University of Minnesota Twin Cities, Minneapolis, MN, USA, in 2025. He also pursued M.S. research at the Indian Institute of Technology Madras, Chennai, India, for one year. He is currently pursuing the Ph.D. degree in electrical engineering at the University of Minnesota Twin Cities. He was a Software Engineer Intern with Cadence Design Systems, San Jose, CA, USA, during the summers of 2023 and 2024. His research interests include analog/RF/mm-wave circuit design automation, circuit sizing and layout synthesis, and machine learning and generative AI techniques for electronic design automation.

Ramesh Harjani

Ramesh Harjani received the B.S. degree from the Birla Institute of Technology and Science, Pilani, India, in 1982, the M.S. degree from the Indian Institute of Technology Delhi, India, in 1984, and the Ph.D. degree from Carnegie Mellon University, Pittsburgh, PA, USA, in 1989, all in electrical engineering. He is currently the Edgar F. Johnson Professor in the Department of Electrical and Computer Engineering at the University of Minnesota. Previously, he worked at Mentor Graphics Corporation, San Jose, CA, USA, and co-founded Bermai, Inc., a wireless multimedia CMOS chip startup, in 2001. He has also been a Visiting Professor at Lucent Bell Labs and the Army Research Laboratory. His research interests include analog/RF circuits for wired and wireless communications. He is a Fellow of the IEEE.

Sachin S. Sapatnekar

Sachin S. Sapatnekar received the B.Tech. degree from the Indian Institute of Technology Bombay, India, the M.S. degree from Syracuse University, NY, USA, and the Ph.D. degree from the University of Illinois at Urbana-Champaign, IL, USA. He is currently a Distinguished McKnight University Professor and holds the Robert and Marjorie Henle Chair at the University of Minnesota. His research interests include analog and digital design automation, machine learning for EDA, and emerging computing technologies. He has served as General Chair and Technical Program Chair of DAC, ISPD, and the Tau Timing Workshop, and as Editor-in-Chief of IEEE TCAD from 2010 to 2013. His honors include 12 Best Paper Awards, two ICCAD 10-Year Retrospective Most Influential Paper Awards, the SRC Technical Excellence Award (2003), the SIA University Research Award (2013), and the ACM SIGDA Pioneering Achievement Award (2026). He is a Fellow of the ACM and IEEE.

References

  1. [1]J. R. Koza, F. H. Bennett, D. Andre, and M. A. Keane, “Automated design of both the topology and sizing of analog electrical circuits using genetic programming,” in Artificial Intelligence in Design ’96. Dordrecht, Netherlands: Springer, 1996, pp. 151–170.
  2. [2]W. Kruiskamp and D. Leenaerts, “DARWIN: CMOS op-amp synthesis by means of a genetic algorithm,” in Proceedings of the ACM/IEEE Design Automation Conference, 1995, pp. 433–438.
  3. [3]B. Liu, Y. Wang, Z. Yu, L. Liu, M. Li, Z. Wang, J. Lu, and F. V. Fernández, “Analog circuit optimization system based on hybrid evolutionary algorithms,” Integration, vol. 42, no. 2, pp. 137–148, 2009.DOI
  4. [4]G. Gielen, H. Walscharts, and W. Sansen, “Analog circuit design optimization based on symbolic simulation and simulated annealing,” IEEE Journal of Solid-State Circuits, vol. 25, no. 3, pp. 707–713, Jun. 1990.
  5. [5]E. S. Ochotta, R. A. Rutenbar, and L. R. Carley, “Synthesis of high performance analog circuits in ASTRXOBLIX,” IEEE Transactions on Computer-Aided Design of Integrated Circuits and Systems, pp. 273–294, Mar. 1996.
  6. [6]R. A. Vural and T. Yildirim, “Analog circuit sizing via swarm intelligence,” AEU – International Journal of Electronics and Communications, vol. 66, no. 9, pp. 732–740, Sep. 2012.DOI
  7. [7]B. Liu, F. V. Fernández, G. E. Gielen, R. Castro-López, and E. Roca, “A memetic approach to the automatic design of high-performance analog integrated circuits,” ACM Transactions on Design Automation of Electronic Systems, vol. 14, no. 3, pp. 1–24, May 2009.DOI
  8. [8]N. Lourenço and H. C. Horta, “HGENO-MOF: multi-objective evolutionary synthesis of analog ICs with corners validation,” in Proceedings of Genetic and Evolutionary Computation, Jul. 2012, pp. 1119–1126.DOI
  9. [9]W. Lyu, F. Yang, C. Yan, D. Zhou, and X. Zeng, “Batch Bayesian optimization via multi-objective acquisition ensemble for automated analog circuit design,” in Proceedings of the International Conference on Machine Learning, Jul. 2018, pp. 3306–3314.
  10. [10]W. Lyu, P. Xue, F. Yang, C. Yan, Z. Hong, X. Zeng, and D. Zhou, “An efficient Batch Bayesian optimization approach for automated optimization of analog circuits,” IEEE Transactions on Circuits and Systems I, vol. 65, no. 6, pp. 1954–1967, Jun. 2018.DOI
  11. [11]S. Zhang, F. Yang, D. Zhou, and X. Zeng, “An efficient asynchronous batch Bayesian optimization approach for analog circuit synthesis,” in Proceedings of the ACM/IEEE Design Automation Conference, Jul. 2020, 6 pages.DOI
  12. [12]K. Settaluri, A. Haj-Ali, Q. Huang, K. Hakhmanesh, and B. Nikolic, “AutoCkt: Deep reinforcement learning of analog circuit designs,” in Proceedings of the Design, Automation & Test in Europe Conference, Mar. 2020, 6 pages.arxiv.org/abs/2001.01808
  13. [13]Z. Zhao and L. Zhang, “Deep reinforcement learning for analog circuit sizing,” in Proceedings of the IEEE International Symposium on Circuits and Systems, Oct. 2020, 5 pages.DOI
  14. [14]M. Choi, Y. Choi, K. Lee, and S. Kang, “Reinforcement learning-based analog circuit optimizer using $m/I_D$ for sizing,” in Proceedings of the ACM/IEEE Design Automation Conference, Jul. 2023, 6 pages.DOI
  15. [15]Y. Yin, Y. Wang, B. Xu, and P. Li, “ADO-LLM: Analog design Bayesian optimization with in-context learning of large language models,” in Proceedings of the IEEE/ACM International Conference on Computer-Aided Design, Oct. 2024, 9 pages.
  16. [16]Y. Lai, S. Lee, G. Chen, S. Poddar, M. Hu, D. Z. Pan, and P. Luo, “AnalogCoder: Analog circuit design via training-free code generation,” in Proceedings of the AAAI Conference on Artificial Intelligence, Feb. 2025, 9 pages.
  17. [17]J. Lee and J. Kim, “Efficient global optimization of analog circuits using predictive response surface models on discretized design space,” IEEE Design & Test, vol. 33, no. 5, pp. 16–27, 2016.
  18. [18]Y. Li, Y.-L. Li, and S.-M. Yu, “Design optimization of a current mirror amplifier integrated circuit using a computational statistics technique,” Mathematics and Computers in Simulation, vol. 79, no. 4, pp. 1165–1177, 2008.DOI
  19. [19]A. F. Budak, P. Bhansali, B. Liu, N. Sun, D. Z. Pan, and C. V. Kashyap, “DNN-Opt: An RL inspired optimization for analog circuit sizing using deep neural networks,” in Proceedings of the ACM/IEEE Design Automation Conference, Dec. 2021, pp. 1219–1224.arxiv.org/abs/2110.00211
  20. [20]S. Ghosh, E. Y. Gebru, C. V. Kashyap, R. Harjani, and S. S. Sapatnekar, “Accelerating OTA circuit design: Transistor sizing based on a transformer model and precomputed lookup tables,” in Proceedings of the Design, Automation & Test in Europe Conference, Apr. 2025, 7 pages.arxiv.org/abs/2502.03605
  21. [21]H. Wang, K. Wang, J. Yang, L. Shen, N. Sun, H.-S. Lee, and S. Han, “GCN-RL circuit designer: Transferable transistor sizing with graph neural networks and reinforcement learning,” in Proceedings of the ACM/IEEE Design Automation Conference, Jul. 2020, 6 pages.DOI
  22. [22]I. Abel, M. Neuner, and H. E. Graeb, “A hierarchical performance equation library for basic op-amp design,” IEEE Transactions on Computer-Aided Design of Integrated Circuits and Systems, vol. 41, no. 7, pp. 1976–1989, Jul. 2022.DOI
  23. [23]H. Graeb, S. Zizala, J. Eckmueller, and K. Antreich, “The sizing rules method for analog integrated circuit design,” in Proceedings of the IEEE/ACM International Conference on Computer-Aided Design, Nov. 2001, pp. 343–349.
  24. [24]T. Maass, H. Graeb, and U. Schlichtmann, “The sizing rules method for CMOS and bipolar analog integrated circuit synthesis,” IEEE Transactions on Computer-Aided Design of Integrated Circuits and Systems, vol. 27, no. 12, pp. 2209–2222, Dec. 2008.
  25. [25]R. Harjani, R. Rutenbar, and L. Carley, “OASYS: A framework for analog circuit synthesis,” IEEE Transactions on Computer-Aided Design of Integrated Circuits and Systems, vol. 8, no. 12, pp. 1247–1266, Dec. 1989.
  26. [26]H. Koh, C. Sequin, and P. Gray, “OPASYN: A compiler for CMOS operational amplifiers,” IEEE Transactions on Computer-Aided Design of Integrated Circuits and Systems, vol. 9, no. 2, pp. 113–125, Feb. 1990.
  27. [27]E. Y. Gebru, S. Ghosh, R. Harjani, and S. S. Sapatnekar, “Parasitics-aware framework for integrated OTA sizing and layout synthesis in FinFET technologies,” in Proceedings of the IEEE International Symposium on Quality Electronic Design, Apr. 2026, 9 pages.DOI
  28. [28]W. Daems, G. Gielen, and W. Sansen, “Simulation-based generation of posynomial performance models for the sizing of analog integrated circuits,” IEEE Transactions on Computer-Aided Design of Integrated Circuits and Systems, vol. 22, no. 5, pp. 517–534, May 2003.
  29. [29]P. Mandal and V. Visvanathan, “CMOS op-amp sizing using a geometric programming formulation,” IEEE Transactions on Computer-Aided Design of Integrated Circuits and Systems, vol. 20, no. 1, pp. 22–38, Jan. 2001.
  30. [30]M. del Mar Hershenson, S. P. Boyd, and T. H. Lee, “GPCAD: A tool for CMOS op-amp synthesis,” in Proceedings of the IEEE/ACM International Conference on Computer-Aided Design, Nov. 1998, pp. 296–303.
  31. [31]S. Ghosh, R. Harjani, and S. S. Sapatnekar, “NEMESIS: Netlist-driven modeling and equation synthesis with inversion-aware SPICE anchoring,” in Proceedings of the IEEE International Conference on LLM-Aided Design, Aug. 2026, 7 pages.arxiv.org/abs/2607.05657
  32. [32]F. Silveira, D. Flandre, and P. Jespers, “A gm/ID based methodology for the design of CMOS analog circuits and its application to the synthesis of a silicon-on-insulator micropower OTA,” IEEE Journal of Solid-State Circuits, vol. 31, no. 9, pp. 1314 – 1319, Oct. 1996.DOI
  33. [33]P. G. A. Jespers, The gm/ID Methodology, A Sizing Tool for Low-voltage Analog CMOS Circuits: The Semi-empirical and Compact Model Approaches. New York, USA: Springer, 2009.DOI
  34. [34]P. Jespers and B. Murmann, Systematic Design of Analog CMOS Circuits: Using Pre-Computed Lookup Tables. Cambridge, UK: Cambridge University Press, 2017.DOI
  35. [35]P. Lewis, E. Perez, A. Piktus, F. Petroni, V. Karpukhin, N. Goyal, H. Küttler, M. Lewis, W.-t. Yih, T. Rocktäschel, S. Riedel, and D. Kiela, “Retrieval-augmented generation for knowledge-intensive NLP tasks,” in Advances in Neural Information Processing Systems, 2020, 16 pages.arxiv.org/abs/2005.11401
  36. [36]J. H. Holland, Adaptation in Natural and Artificial Systems: An Introductory Analysis with Applications to Biology, Control, and Artificial Intelligence. Ann Arbor, Michigan, USA: University of Michigan Press, 1975.
  37. [37]J. Kennedy and R. Eberhart, “Particle swarm optimization,” in Proceedings of the International Conference on Neural Networks, vol. 4, Dec. 1995, pp. 1942–1948.
  38. [38]R. Storn and K. Price, “Differential evolution: A simple and efficient heuristic for global optimization over continuous spaces,” Journal of Global Optimization, vol. 11, no. 4, pp. 341–359, Dec. 1997.DOI
  39. [39]N. Shervashidze, P. Schweitzer, E. J. van Leeuwen, K. Mehlhorn, and K. M. Borgwardt, “Weisfeiler-Lehman graph kernels,” Journal of Machine Learning Research, vol. 12, pp. 2539–2561, Nov. 2011.
  40. [40]P. Jaccard, “Étude comparative de la distribution florale dans une portion des alpes et des jura,” Bulletin de la Société Vaudoise des Sciences Naturelles, vol. 37, pp. 547–579, Jan. 1901.DOI
  41. [41]P. E. Allen and D. R. Holberg, CMOS Analog Circuit Design, 3rd ed. New York, NY, USA: Oxford University Press, 2011.
  42. [42]ALIGN-nanoLayout, “ALIGN: Analog Layout, Intelligently Generated from Netlists,” https://github.com/ALIGN-analoglayout/ALIGN-public/tree/master/CircuitsDatabase/Generalized_Testbenches, 2026, accessed on Jun. 01, 2026.
  43. [43]OpenAI, “GPT-5.5 model,” https://openai.com/index/introducing-gpt-5-5/, 2026, accessed on May 01, 2026.
  44. [44]T. J. Santner, B. J. Williams, and W. I. Notz, The Design and Analysis of Computer Experiments, 2nd ed. New York, USA: Springer, 2019.

Paper details

vlnc:W_q5jg9s33
Version
v1 · latest
Published
Fresh — added to the Hub within the last 30 days
Added
License
CC BY 4.0
Usage

views · 2

downloads · 0

A view counts once per browser per day, a download each time the PDF is requested.

  • Computer Science

OpenAI ChatGPT

View statement
Looking for related papers…
Contents