Abstract
This article describes a tool for mining assertions as cause-effect relations over real-time from traces of hybrid systems. The tool mines sequence expressions that describe the qualitative (ordering of events) and quantitative (temporal distance between events) relationships between predicates over real variables (PORVs) for a hybrid system. Given a target predicate, the miner learns candidate sequence expressions that are likely to be the cause of the target's truth. The tool uses interval abstractions to represent the real-time truth of predicates and uses interval arithmetic to reason about the relationships between them. Expressing cause-effect relations in a hybrid system is a challenge due to the intricate ways in which system artefacts interact with one another. Additionally, expressing such interactions over time is a tedious error-prone task for a verification engineer, usually involving large amounts of manual effort, and most often incomplete. Results are demonstrated on synthetic traces as well as traces from simulations of industrial designs.
Introduction
In semiconductor design, hybrid systems—which integrate discrete digital logic with continuous analog signals—present a significant verification challenge. While digital systems operate on discrete clock cycles, analog components function in "dense time," characterized by continuous and interdependent behaviors. Engineers typically use formal assertions (logical statements about expected behavior) to validate these systems, but manually crafting these rules for complex analog circuits is a tedious, error-prone process. AMS-Miner$^T$ addresses this by automatically discovering logical "cause-effect" assertions directly from simulation traces.
The AMS-Miner$^T$ Framework
The tool employs a multi-stage pipeline to transform raw simulation data into human-readable specifications.
The framework operates in four primary stages:
- Booleanization: Continuous, real-valued traces are converted into "truth intervals." For instance, a voltage signal is mapped to time intervals where a specific predicate (e.g., $v > 4.2V$) holds true.
- Interval Abstraction: The tool reasons about these intervals using Minkowski arithmetic, allowing it to manage the inherent "fuzziness" of real-time delays without the state explosion common in traditional formal verification.
- Decision Tree Learning: Using algorithms inspired by ID3/C4.5, the miner constructs a decision tree to identify which preceding events (causes) most accurately predict a target event (effect).
- Temporal Sequencing: By shifting truth intervals backward in time, the tool identifies causal links across different time windows and calculates tight delay bounds using Minkowski differences.
From Analog Traces to Logical Predicates
A core innovation of AMS-Miner$^T$ is its ability to abstract complex analog behavior into discrete logical intervals. This process is essential for handling the continuous nature of analog signals within a formal logic framework.
As shown in the figure above, raw signals are evaluated against thresholds (like $0.9 \times V_r$) to generate boolean predicates. This abstraction allows the mining process to be extremely efficient, processing hundreds of thousands of data points in seconds.
Evaluation and Key Results
The tool was evaluated on both synthetic datasets and an industrial Low-Dropout (LDO) voltage regulator netlist.
- Efficiency: For an industrial LDO trace with over 500,000 data points, Booleanization took less than 3 seconds, and the assertion mining itself completed in under one second.
- Coverage: On synthetic data, the tool achieved up to 100% trace coverage, successfully finding assertions to explain every instance of target predicates even in chaotic traces.
- Discovery: In the LDO design, AMS-Miner$^T$ discovered both expected behaviors and previously undocumented relationships, such as precise timing links between different voltage rise thresholds.
Conclusion and Future Work
AMS-Miner$^T$ demonstrates that treating time as a continuum and using interval-based abstractions can significantly streamline the verification of hybrid systems. While the tool currently requires expert-provided predicates, future work aims to automate predicate discovery. By reducing the manual effort required for assertion generation, AMS-Miner$^T$ provides a scalable path toward more robust analog and mixed-signal design verification.
Summary updated