AutoICE: Automated Solutions Across Domains

• AutoICE is a multi-domain framework that applies automated, LLM-enhanced, and specialized methods to challenges in code verification, deicing, medical imaging, navigation, and astrochemical analysis.
• Its methodologies leverage evolutionary search, MEMS vibroacoustic sensing, robotic manipulation, trajectory optimization, and deep neural inference to achieve superior performance compared to traditional baselines.
• Across applications, AutoICE systems demonstrate significant improvements such as increased code verification rates, energy-efficient deicing, precise catheter repositioning, reduced collision impacts in navigation, and accurate ice composition estimation.

AutoICE refers to a set of independently developed methodologies, algorithms, and systems sharing the prefix "AutoICE," whose technical and application domains are distinct: (1) automated synthesis of verifiable C code using LLM-driven evolutionary search (Luo et al., 8 Dec 2025), (2) vibroacoustic sensing and deicing in microacoustic wave devices (Karimzadeh et al., 2023), (3) robotic repositioning of ICE catheters for view recovery in cardiac imaging (Kim et al., 2022), (4) autonomous surface vessel navigation in ice fields (Schaetzen et al., 2024), and (5) automated estimation of astronomical ice compositions via IR spectroscopy and neural inference (Megías et al., 4 Sep 2025). Each "AutoICE" is characterized by task-specific technical architectures and evaluation metrics rooted in its respective field. The following sections review these systems individually, focusing on their technical frameworks and empirical performance.

1. LLM-Driven Evolutionary Synthesis of Verifiable C Code

AutoICE as described in "AutoICE: Automatically Synthesizing Verifiable C Code via LLM-driven Evolution" (Luo et al., 8 Dec 2025) is a code synthesis and verification framework that integrates LLMs with classical evolutionary algorithms. Its central goal is to generate C code with ACSL (ANSI/ISO C Specification Language) annotations from natural-language requirements and ensure formal correctness via static verification.

System Architecture

At its core, AutoICE models each solution candidate as a tuple $$I = (\text{code\_with\_ACSL}, \text{base\_pass}, \text{wp\_pass}, \text{base\_info}, \text{wp\_info})$$, where base_pass and wp_pass are binary indicators of successful parsing and proof obligation discharge through Frama-C and its WP plugin, respectively. Fitness is defined as $$\text{fitness}(I) = \text{base\_pass} + \text{wp\_pass}$$.

The evolutionary loop comprises:

• Diverse individual initialization: Syntactically and semantically distinct initial candidates are produced by varying both the code-generation and specification-generation prompt strategies (e.g., chain-of-thought vs. step-back prompting).
• Collaborative crossover: For parent selection, two high-fitness individuals are chosen, and the LLM is instructed (with code and verification reports as context) to produce offspring infusing strengths of both.
• Self-reflective mutation: The LLM is prompted to diagnose and repair failures (through analysis of parsed feedback) to discover novel corrections or specification insights.

Algorithmically, the method is summarized as follows:

function InitializePopulationLLM(R, S): pop ← {} for i in 1…S: strategy ← pick one reasoning style code_i ← LLM(prompt_for_phase1(R, strategy)) spec_i ← LLM(prompt_for_phase2(R, code_i, strategy)) I_i ← (full_code=spec_i+code_i; evaluate base_pass, wp_pass) pop ← pop ∪ {I_i} return pop

Crossover and mutation are similarly LLM-mediated, with verification feedback influencing prompts at each stage.

Verification Toolchain and Termination

AutoICE's pipeline connects LLM generation with formal method back-ends:

• Frama-C parses and checks ACSL-annotated C code,
• WP plugin generates proof obligations,
• Why3 translates these to first-order logic,
• SMT solvers (Alt-Ergo, CVC4, Z3) discharge the obligations, with failure/success information relayed to the LLM-driven evolutionary operators.

The search halts upon successful generation (fitness = 2) or after reaching a fixed generation cap.

Empirical Results

Testing across several datasets and eight LLM backbones (7B–671B parameter scales), AutoICE achieved high one-shot verification rates:

Dataset	SOTA baseline	AutoICE
FM (full spec inputs)	85.36%	90.36%
FM_dev (full spec inputs)	80.00%	86.67%
FM (developer-friendly)	65.00%	88.33%
FM_dev (developer-friendly)	68.33%	83.33%

AutoICE consistently outperformed zero-shot/fine-tuned LLM and single-agent refinement baselines, with a notable ability to discover and formalize implicit requirements from unconstrained functional descriptions (Luo et al., 8 Dec 2025).

2. Vibroacoustic Sensing and On-Demand Deicing via Electrode Switching

In the domain of microacoustic sensing and actuation, AutoICE (Karimzadeh et al., 2023) designates a chip-scale system that leverages dual-mode (Rayleigh standing surface acoustic wave, sSAW; and thickness-shear plate wave, PW) operation for precision temperature sensing, phase change detection, and actively powered deicing.

Physical Principles

Two elastic modes are exploited:

• sSAW: Launched by anti-phase driving of interdigital transducers (IDTs), with resonance governed by $f_n = n v_R / 2L$. Here, $v_R$ is Rayleigh wave speed, $L$ the cavity length.
• PW: Excited by in-phase wiring of the same IDTs via relays, producing a shear-horizontal mode with resonance $f_m = m v_S / (2h)$, $v_S$ the shear speed, $h$ plate thickness.

Sensing and Deicing Operation

Fast electronic switching (relay and microcontroller logic) reconfigures electrodes, toggling between high-sensitivity, low-power PW sensing (temperature and ice detection, <1mW) and high-power sSAW deicing (∼7W for 10s).

Upon detection of a sharp attenuation in reflection coefficient (|S11| drop by ~4–6 dB upon icing), the system automatically activates sSAW for brief, energy-efficient deicing.

Energy and Performance Metrics

• Melting a 0.28g ice droplet required ~70–80J acoustic/electrical input; by contrast, a resistive glass heater would consume ~900J for the same task.
• Measured temperature sensitivity is k_T ≈ -1.7 kHz/K with ±0.25K calibration accuracy.
• Full relay sensing-actuation cycle achieved detection latencies below 2s (Karimzadeh et al., 2023).

Integration and Scalability

Frequency-based sensing (PW resonance) is invariant under moderate surface loading or functionalization, and the architecture is compatible with wafer-scale calibration and integration into lab-on-chip, IoT, or large-area anti-icing arrays.

3. Automated ICE Catheter Tip Repositioning for Cardiac Imaging

AutoICE (Kim et al., 2022), in the context of intra-cardiac echocardiography (ICE), denotes a robotic system for precise, reproducible catheter tip repositioning and automated view recovery in clinical interventions.

Robotic System Design

The system uses a four-DoF tendon-driven manipulator directly interfacing with a commercial ICE catheter. Joint-space coordinates $q = [\varphi_1, \varphi_2, \varphi_3, d_4]$ represent two-plane bending, bulk handle rotation, and axial insertion.

• Online, the system builds a joint-space roadmap of visited configurations, with user-bookmarked anatomic views stored as target poses.
• Upon request, A* path planning returns optimal joint sequences for view recovery.

Kinematics and Control

Closed-form direct kinematics derive the tip pose. Inverse kinematics use a decoupled numerical scheme due to underactuation and tendon-coupled curvature, with damped Gauss–Newton updates for convergence.

• Position and orientation control run via PID at the joint level, with hysteresis compensation.

Empirical Validation

• In vitro (phantom), mean tip position error: 0.67 ± 0.79 mm; orientation error: 0.37 ± 0.19°.
• In vivo (swine), mean position error: 2.09 ± 0.90 mm; orientation: 3.93 ± 2.07°.
• The system was successfully applied to fluoroscopy-free transeptal puncture, supporting clinical applications with reduced operator burden and radiation exposure (Kim et al., 2022).

4. Autonomous Surface Ship Navigation in Broken Ice (AUTO-IceNav Framework)

While distinct in spelling, AUTO-IceNav (Schaetzen et al., 2024) employs principles and architecture relevant to the general "AutoICE" schema, specifically in the context of robotic autonomy in hazardous natural environments.

Planning Architecture

The navigation framework utilizes:

• Receding-horizon trajectory optimization: At each replanning step, the system updates an ice-floe segmentation map and computes a safe, energy-efficient trajectory.
• Lattice-based path planning: Motion primitives are generated for a unicycle-approximated vessel, with A* search over primitives accounting for ice collision penalties.
• Hybrid cost function: Primary path cost penalizes both path length and the predicted kinetic energy loss from modeled ship–ice impacts, using a physics-based collision cost function.

Trajectory Refinement

Continuous trajectory optimization (NLP, solved with IPOPT) further refines lattice paths to reduce impact forces while maintaining feasibility within dynamic constraints.

Experimental Evaluation

• In simulation (1000m x 200m channel; 20–50% ice concentration), AUTO-IceNav reduced ship–ice collision kinetic energy loss by up to 59% vs. skeletonization-based baselines, with only ~2% increase in overall path length and transit time.
• In scale-model physical testing, reductions of 52% in work terms vs. skeleton/straight baselines were observed (Schaetzen et al., 2024).

5. Automatic Ice Composition Estimation in Astronomical Infrared Spectra

AICE (Megías et al., 4 Sep 2025), or Automatic Ice Composition Estimator, is a domain-specific pipeline applying artificial neural networks to near- and mid-infrared astronomical spectral analysis, inferring fractional ice composition on interstellar grains.

Data Acquisition and Preprocessing

• Training utilizes 571 laboratory reference spectra (H₂O, CO, CO₂, CH₃OH, NH₃, CH₄ mixes from LIDA/OCdb/NASA-led sources), with baseline subtraction, silicate removal, manual artifact curation, and normalization.
• Input vectorization converts each spectrum to a 3021-dimensional normalized absorbance profile over 2.5–10 µm.

Neural Architecture

• Separate multilayer perceptrons for each target molecule fraction and for temperature estimation.
• Ensemble bagging over 10 random splits provides uncertainty quantification.
• Outputs are renormalized to ensure sum-to-unity constraint and uncertainties propagate via Monte Carlo with ensemble variance.

Performance and Use Cases

• Laboratory mean RMSE per species fraction is ≈2.1–3.3%; temperature RMSE ≈10.8 K.
• JWST application (background stars NIR38, J110621) yields composition estimates consistent with genetic fitting and radiative-transfer models.
• Computational cost is ≲0.1 s per spectrum, permitting near-real-time population-level ice chemistry analyses (Megías et al., 4 Sep 2025).

6. Terminological and Conceptual Distinctions

Despite sharing the "AutoICE" identifier, the reviewed methods are not interrelated in technical content, co-authorship, or application. Each instantiation adopts "AutoICE" as an acronym or prefix specific to its domain—ranging from LLM-accelerated code synthesis, vibroacoustic MEMS sensing, robotic surgical instrumentation, autonomous field robotics, to deep-learning–based astronomical inference.

A plausible implication is that "AutoICE" signals a trend toward integrating automated, data- or model-driven techniques into "ICing"–related or "ICE"–instrumented systems, yet no generic theoretical connection exists between these approaches. Each uses domain-specific architectures, mathematical models, and distinct forms of automation or learning tailored to its problem space.