TRACE: A Multi-Agent System for Autonomous Physical Reasoning for Seismology

Feng Liu; Xin Cui; Jian Xu; Xinghao Wang; Zijie Guo; Jiong Wang; S. Mostafa Mousavi; Xinyu Gu

Figure 1 : TRACE framework for constrained workflow execution and physical interpretation. a , Seismic observations and case-specific geophysical inputs available to TRACE, including station metadata, waveform records, event catalogs, phase-arrival records, fault geometry and velocity models. These inputs define the observational and physical boundary conditions for subsequent analysis. b , Constrained workflow execution. TRACE links waveform preprocessing, event detection and phase picking, association, location and relocation, and seismicity diagnostics through diagnostic checks of intermediate products. Depending on input availability, workflows may enter at different stages. c , Evidence synthesis and physical interpretation. The evidence matrix summarizes qualitative support for candidate interpretations derived from the scientific question and analysis outputs; the lower schematic shows the best-supported interpretation and targets for further validation. d , The coordinated agent layer supports hypothesis planning, workflow execution, diagnostic checking, interpretation evaluation and synthesis. The TRACE record retains data sources, code versions, parameter settings, run metadata and audit logs. Miniature panels illustrate representative inputs, workflow products and evidence domains; quantitative application results are shown in subsequent figures.

Figure 2 : TRACE construction of a reviewable Ridgecrest earthquake catalog workflow. TRACE converts waveform records and task constraints into a documented catalog-building workflow whose data selections, processing parameters, phase picks, associated events, initial locations, local-magnitude estimates and relocation outputs can be inspected. The 2019 Ridgecrest sequence is shown as a worked example: TRACE plans a six-step route and executes waveform acquisition, preprocessing, PhaseNet phase picking, GaMMA association, initial location, and HypoDD-based relocation. Summary cards report 47 stations selected for waveform processing, of which the 30 nearest to the

M_{w}

6.4 epicenter were retained for association and relocation; 34 GiB of requested waveforms; approximately 1.45 million

P

picks and 1.64 million

S

picks; 140,505 associated events with initial locations; 84,474 relocated events; and 4,714 events between the

M_{w}

6.4 and

M_{w}

7.1 earthquakes. These catalog and provenance records provide the observational basis for evaluating triggering pathways from the

M_{w}

6.4 foreshock to the

M_{w}

7.1 mainshock.

Figure 3 : Evidence for delayed cascading activation between the Ridgecrest

M_{w}

6.4 foreshock and

M_{w}

7.1 mainshock. a , Seismicity over 34 h in three windows (0–4, 4–18 and 18–33.8 h) after the foreshock, with mapped SW-NE- and NW-SE-trending surface-rupture traces [ 18 ] , the

M

5.4 analytical separator and inferred activation direction toward the NW-SE fault family. Pale blue and red corridors define domains in c,d ; darker dashed outlines mark the two 5 km comparison regions in e . b , Distances to the nearest mapped surface-rupture trace and directional spectrum; reference directions mark the dominant fault-family orientations, and vertical dashed lines mark 1 and 2 km. c , Hourly seismicity rates in the foreshock- and mainshock-side corridors; shading marks the separator interval. d , Grid-cell activation onset times in the same corridors; horizontal bars denote medians. e , Fixed-window

b

-values for the foreshock and mainshock comparison regions before and after the separator and over the full interval. Error bars denote standard uncertainties,

n

event counts and the gray band the regional background range. f , Qualitative evidence-weighting matrix comparing delayed cascading activation with dynamic triggering, ordinary aftershock decay and a coherent migration front; cells summarize diagnostic consistency and do not represent an independent statistical test.

Figure 4 : Persistent shallow-interface activation before the 2026 Sanriku

M_{J}

7.7 earthquake. The physical synthesis links stage-resolved seismicity, the focused shallow-interface geometry and neighboring historical large-slip regions. a–e , Five-stage evolution of relocated seismicity, from background activity through the November 2025

M_{J}

6.9 compound burst, sustained shallow-interface activation, pre-

M_{J}

7.7 intensification and the

M_{J}

7.7 locked-asperity failure. Dashed outlines mark the

M_{J}

6.9 source region, the broader shallow activation zone and the historically bounded

M_{J}

7.7 patch. The central timeline summarizes the transition from distributed activity around the

M_{J}

6.9 source to the

M_{J}

7.7 rupture. f , Along-profile distance-depth projection of the focused shallow-interface region. The

M_{J}

6.9 and

M_{J}

7.7 reference locations are separated by

\sim

57 km, and

M\geq 4

earthquakes persist in spatially separated patches across the focused profile rather than forming a continuous migration front. g , Three-dimensional synthesis of short-term activation and historical rupture geometry. Historical rupture contours mark neighboring regions of large coseismic slip, whereas the focused shallow-interface region occupies a patch bracketed by these regions, with the later

M_{J}

7.7 rupture occurring within the same bounded patch.

Figure 5 : Seismicity-rate, magnitude–frequency and repeating-earthquake evidence for Sanriku shallow-interface activation. a , Relocated catalog colored by depth in the focused Sanriku region. Stars mark the

M_{J}

6.9 and

M_{J}

7.7 earthquakes (red) and the regional

M

7.5 earthquake (blue). Contours show the 2011

M_{w}

9.0 Tohoku-Oki high-slip region [ 16 ] and the 1994

M_{w}

7.7 Sanriku-Oki and 1968 Tokachi-Oki (approximately

M_{w}

8.3) ruptures [ 24 ] . The inset shows

M\geq 4

events along the

\sim

57-km shallow-interface profile. b , Magnitude-time plots for the November 2025

M_{J}

6.9, December 2025 regional

M

7.5 and April 2026

M_{J}

7.7 sequences, each aligned to its largest-event origin time. Colors denote 0–30, 30–60 and 60–100 km epicentral-distance bins; diamonds mark

M\geq 4

events. c , Long-term seismicity rate, cumulative event count,

b

-value and magnitude of completeness. Lines mark the

M_{J}

6.9 and

M_{J}

7.7 events; shading shows

\pm 1\sigma

uncertainty in

b

-value estimates. d , Repeating-earthquake candidates and high-correlation families near the

M_{J}

6.9 source. Hulls enclose seismicity during the final 2 days before and first 0.5 days after the event; contours show historical coseismic slip. Colors indicate family thresholds; CC, cross-correlation. e , Repeating-family timeline ordered by first event; points denote events, segments family durations and the dashed line the

M_{J}

6.9 origin time. f , Representative waveforms from family F063. Colors denote the two events; vertical lines mark

P

and

S

picks. Labels give station-component cross-correlation and time shift; the title gives the network-median CC.

Figure S1 : Technical architecture and evidence-chain synthesis of TRACE. a , Multi-agent architecture used to convert a seismological request and optional literature or contextual input into candidate plans, executable workflows, scripts, checked data products and a physical interpretation report. Planning, workflow, coding, result-checking, analysis and cross-perspective synthesis agents query the Expert-Validated Case Library, Structured Knowledge Library and Validated Scientific Tools Library. Dashed links distinguish knowledge and case retrieval from tool retrieval and execution. Optional human supervision and expert validation constrain ambiguous task boundaries and return reviewed workflows and summaries to the case library. b , Evidence-chain synthesis used by the Scientific Integration and Inference Module. Checked workflow outputs are converted into perspective-specific evidence summaries, integrated across evidence streams and assembled into a physical interpretation report containing an evidence-supported interpretation, viable alternatives, external validation targets and a trace record.

Levels	Description	Number of Tasks
Level 1: Atomic task capability	Performs fundamental seismological operations with procedural correctness and execution stability, including seismic data acquisition, format conversion, preprocessing, signal feature extraction, and scientific visualization. Emphasizes accurate tool invocation, parameter handling, and computational consistency in routine seismic data processing workflows.	74
Level 2: Multi-step analytical capability	Solves workflow-level seismological problems requiring coordinated multi-tool orchestration and dependency management. Integrates sequential toolchains, heuristic branching strategies, large-scale batch scanning, and parameter-space exploration under physical constraints. Evaluates adaptive pathway selection, intermediate state management, and long-chain analytical coherence.	30
Level 3: Integrated scientific analysis capability	Conducts end-to-end scientific investigations under realistic research scenarios. Integrates cross-module data processing, multi-perspective analysis, hypothesis formulation, and evidence synthesis to assemble evidence-supported, testable interpretations. Assesses workflow coherence, cross-task integration, and interpretation boundaries in complex seismological studies.	2

Table S1 : Task-complexity levels and representative TRACE capabilities in the seismological evaluation framework.

Score	Quantitative Accuracy	Scientific Correctness	Visualization and Reporting Quality
5 points	Numerical outputs match reference solutions within predefined numerical tolerances.	All results are scientifically valid, logically consistent, and fully aligned with task objectives. No conceptual or methodological errors.	Figures are accurate, clearly labeled (axes, units, scales), and publication-ready. Visual design enhances interpretability. Text and figures are fully consistent.
4 points	Minor numerical deviations from reference results but within acceptable tolerance. No impact on overall conclusions.	Results are scientifically correct with only minor interpretative omissions or limited analytical depth.	Figures are correct and interpretable, but minor issues exist (e.g., formatting inconsistencies, limited annotations). Overall presentation is clear.
3 points	Moderate deviation from reference values, but core trends or magnitudes remain correct.	Partial correctness: main scientific conclusion is identifiable, but secondary errors or incomplete interpretation exist.	Figures are understandable but contain noticeable labeling issues, limited clarity, or insufficient explanation.
2 points	Significant numerical errors or unstable results that affect reliability.	Results contain conceptual misunderstandings or incorrect scientific interpretations.	Figures are poorly constructed or partially incorrect, limiting interpretability.
1 point	Numerical results are incorrect or fail to match reference benchmarks.	Scientific conclusions are invalid or inconsistent with task requirements.	No meaningful figures or coherent result presentation provided.

Table S2 : Five-point criteria for qualitative Level 1 and Level 2 benchmark outputs.

Provenance layer	Information retained	Audit function
Input boundary	Scientific question; data sources; spatial and temporal selections; permitted contextual information; stage at which external evidence entered	Establishes what information was available at each stage
Planning	Physical constraints; candidate physical interpretations; selected and retained analysis branches	Links the scientific question to the analyses without prescribing a conclusion
Execution	Tool choices; executable scripts; parameter settings; input–output paths; intermediate products	Reconstructs the computational path from observations to analytical products
Diagnostics and revision	Computational and physical checks; failed steps; diagnostic flags; permitted revisions; intervention status and source, when applicable	Distinguishes accepted outputs from unresolved failures and records how the workflow changed
Evidence extraction	Source data product; method and parameters; diagnostic status; interpretation-relevant observation	Links each evidence statement to a checked analytical output
Synthesis	Supporting, inconsistent or unresolved relations; evidence-supported interpretation; viable alternatives; external validation targets	Links the physical interpretation to the evidence while preserving its boundaries

Table S3 : Core fields in the curated TRACE provenance record. Each workflow layer retains the information needed to trace a physical interpretation back to its observational and computational basis.

Benchmark group	Benchmark category	Capability tested	Operation pattern	Tasks (n)
L1-A	Data access and retrieval	Data retrieval	Single-step	7
L1-B	Data representation and formatting	Data formatting	Single-step	12
L1-C	Signal processing	Signal processing	Single-step	22
L1-D	Statistics and feature extraction	Statistical feature extraction	Single-step	18
L1-E	Scientific visualization	Scientific plotting	Single-step	15
Level 1 subtotal				74

Table S4 : Coverage of Level 1 and Level 2 benchmark tasks. Counts are grouped by benchmark group, tested capability and operation pattern.

Task ID	Benchmark category	Representative task	Operation pattern	Complexity
L1-A-1	Data access and retrieval	Download an event catalog from an FDSN server	Single-step	Low
L1-B-1	Data representation and formatting	Convert ASCII waveform data to MiniSEED	Single-step	Low
L1-C-1	Signal processing	Apply an AIC picker to a seismic trace	Single-step	Low
L1-D-1	Statistics and feature extraction	Compute an array response function	Single-step	Low
L1-E-1	Scientific visualization	Plot a focal-mechanism beachball from a QuakeML event	Single-step	Low
L2-A-1	Sequential chaining	Estimate back azimuth from STEAD waveforms using BazNet	Multi-step	Medium

Table S5 : Representative Level 1 and Level 2 benchmark tasks. One task is shown for each benchmark category.

Model	Model–task runs	Success rate	Statistical analysis rating	Visualization rating	Mean correction rounds	Mean score
Claude Sonnet 4	104	98.08%	4.49	4.88	2.28	4.79
Gemini 3 Flash preview	104	100.00%	4.64	4.86	2.26	4.82
GPT-5.1	104	100.00%	4.72	4.83	1.64	4.85

Table S6 : Model-level results for the Level 1 and Level 2 benchmark. Each backbone contributed 104 model–task runs. Ratings and mean scores are reported on a five-point scale. Statistical-analysis and visualization ratings were calculated only when the corresponding component formed part of a task’s predefined output requirements.

Task ID	Case study	Scientific question	Workflow inputs and information boundary	Required report
L3-001	Ridgecrest	How was seismicity organized between the $M_{w}$ 6.4 foreshock and the $M_{w}$ 7.1 mainshock?	Initial observations: continuous waveforms and station metadata; auxiliary geospatial data: mapped surface-rupture traces; quality control and alignment: reference catalogs	Evidence-supported physical interpretation and viable alternatives, followed by retrospective comparison with published interpretations after the analysis was complete
L3-002	Sanriku	What evidence constrains shallow-interface activation before the April 2026 $M_{J}$ 7.7 rupture?	Initial observations: JMA event catalog and phase-arrival records; analysis-stage data: event waveforms for repeating-earthquake analysis; synthesis-stage constraints: historical rupture models and convergence estimates; external validation: independent S-net strain observations	Evidence-supported physical interpretation, viable alternatives and external validation targets

Table S7 : Integrated Level 3 scientific case-study evaluations. The information boundary distinguishes initial observations from auxiliary, quality-control, synthesis and external-validation inputs.

Dimension	Assessment criterion	Common failure mode	Assessment evidence
Planning	Scientific question decomposed into executable, physically relevant analyses	Missing analysis branch or unsupported assumption	Workflow plan and task decomposition
Implementation	Executable scripts and tool calls linked to expected products	Broken execution, unstable parameters or missing quality control	Code, logs and output products
Synthesis	Report separates observations from interpretation and records viable alternatives and external validation targets	Overclaiming, unsupported interpretation or unacknowledged uncertainty	Structured report and expert review

Table S8 : Expert-assessment dimensions for Level 3 case reports. The table defines the review criteria, common failure modes and the records used for assessment.

Processing output	Count (n)	Workflow use
Selected three-component stations	47	Waveform processing and phase picking station set
Stations used for association and relocation	30	Stations nearest to the $M_{w}$ 6.4 epicenter
Quality-controlled station-day waveform inputs	853	Waveform inputs retained after station-day quality control
$P$ arrivals retained after PhaseNet filtering	1,454,311	Thresholded PhaseNet $P$ picks before 30-station association
$S$ arrivals retained after PhaseNet filtering	1,641,210	Thresholded PhaseNet $S$ picks before 30-station association
Picks assigned to associated events	2,120,090	Event-level picks after station and association filtering

Table S9 : Processing counts for the Ridgecrest TRACE catalog.

Stage	Setting	Value
Phase picking	PhaseNet probability threshold	$P\geq 0.4$ ; $S\geq 0.4$
Phase picking	Waveform sampling rate	100 Hz
Association	Association model	GaMMA Bayesian Gaussian-mixture model
Association	DBSCAN pre-clustering	Enabled; $\epsilon=10$ km; minimum samples, 3
Association	Oversampling factor	3
Association	Amplitude term	Not used

Table S10 : Association and relocation parameters for the Ridgecrest catalog.

Catalog	Events	$N(M\geq 1)$	$N(M\geq 2)$	$N(M\geq 3)$	Median $M$	Use in TRACE audit
Ross et al. high-resolution catalog, $n_{\mathrm{branch}}>1$ [ 35 ]	4,138	2,499	605	111	1.15	High-resolution reference
USGS/ANSS ComCat (USGS/CI) [ 45 ]	3,484	2,686	704	145	1.37	Operational reference
TRACE associated catalog before relocation	6,351	3,907	817	163	1.14	Unrelocated TRACE reference
TRACE HypoDD relocated catalog	4,714	3,205	734	151	1.23	Final catalog used for diagnostic tests

Table S11 : Reference-catalog comparison for the Ridgecrest

M_{w}

6.4 to

M_{w}

7.1 interval.

Figure S3 : Relocation sharpening for the Ridgecrest

M_{w}

6.4 to

M_{w}

7.1 interval. a , Initial associated events and double-difference relocated events between the

M_{w}

6.4 and

M_{w}

7.1 earthquakes. Blue points show the GaMMA-associated catalog, red points show relocated events, and stars mark the two earthquakes. Relocation retains the principal activation corridor while reducing diffuse scatter away from the corridor. b , Cumulative distribution of absolute cross-profile distance from the activation corridor for the initial and relocated catalogs. The relocated catalog is more concentrated near the corridor. c , Along-profile and cross-profile density distributions for the same events. Relocation preserves the broad along-profile activation pattern while sharpening the cross-profile distribution. These diagnostics evaluate whether the relocated catalog is suitable for the spatial analyses in Supplementary Note 4. They are not used by themselves as evidence for delayed cascading activation.

Figure S4 : Spatial coverage of Ridgecrest reference catalogs and the

M_{w}

6.4 to

M_{w}

7.1 analysis subset. a–c , Full-window spatial distributions for the Ross et al. high-resolution catalog [ 35 ] , the USGS/ANSS ComCat operational catalog [ 45 ] and the TRACE relocated catalog from 4 to 26 July 2019. The black rectangle marks the analysis region between the

M_{w}

6.4 and

M_{w}

7.1 earthquakes, and stars mark their epicenters. Event counts report the number of events shown under the corresponding catalog selection. d–f , Enlarged spatial distributions between the

M_{w}

6.4 and

M_{w}

7.1 origin times. The three catalogs recover the same fault-aligned activation corridor, supporting use of the TRACE relocated catalog for delayed cascading diagnostics.

Figure S5 : Temporal and magnitude consistency of Ridgecrest reference catalogs. a , Full-window time–magnitude distributions for the Ross et al. high-resolution catalog [ 35 ] , the USGS/ANSS ComCat operational catalog [ 45 ] and the TRACE relocated catalog from 4 to 26 July 2019. The yellow and red dashed lines mark the

M_{w}

6.4 and

M_{w}

7.1 origin times, respectively, and the shaded interval marks the analysis window used for delayed cascading diagnostics. Event counts report the full-window catalog selections shown in this panel. b , Cumulative magnitude-frequency distributions for events in the analysis window. c , Event counts above

M\geq 1

,

M\geq 2

and

M\geq 3

in the same window, matching the magnitude thresholds used to compare catalog completeness for the diagnostic analyses. d , Hourly and cumulative event counts after the

M_{w}

6.4 earthquake. Thin lines show hourly counts and thick lines show cumulative counts. These comparisons show that the TRACE relocated catalog preserves the main temporal and magnitude structure of independent catalogs over the interval used for delayed cascading diagnostics.

Figure S6 : Spatiotemporal activation during the Ridgecrest

M_{w}

6.4 to

M_{w}

7.1 interval. a–d , Relocated seismicity in successive time windows after the

M_{w}

6.4 earthquake and over the full interval between the

M_{w}

6.4 and

M_{w}

7.1 earthquakes. The windows contain 550, 1,969, 2,194 and 4,714 events, respectively. Points are colored by time after the

M_{w}

6.4 earthquake; mapped surface-rupture traces [ 18 ] and earthquake locations provide spatial reference. e , Cumulative activation of 1 km grid cells and newly activated cells through time, with the shaded band marking the

M

5.4 analytical separator. f , Spatial distribution of grid cell activation onset times. g , Distribution of grid cell onset times, with a median of 6.2 h. h , Comparison of onset times between the

M_{w}

6.4 and

M_{w}

7.1 side corridors. i , Time-distance structure in easting and northing components. j , Hourly centroid distance to the

M_{w}

7.1 epicenter and event counts. The panels show delayed, spatially organized activation toward the

M_{w}

7.1 fault system. Timing alone is not used by itself to determine the triggering mechanism.

Figure S7 : Organization of Ridgecrest seismicity near mapped surface ruptures. a,b , Relocated seismicity during the

M_{w}

6.4 to

M

5.4 and

M

5.4 to

M_{w}

7.1 intervals, plotted with mapped surface-rupture traces [ 18 ] and the

M_{w}

6.4,

M

5.4 and

M_{w}

7.1 locations. The windows contain 2,420 and 2,208 events, respectively. c , Cumulative distribution of distance to the nearest mapped surface-rupture trace, with 1 km and 2 km references. d , Directional

L

strength through time, with the SW–NE and NW–SE fault-family directions marked. e , Integrated directional spectrum over the interval between the

M_{w}

6.4 and

M_{w}

7.1 earthquakes, showing enhanced directional strength near 48 ∘ and 128 ∘ . f , Cumulative activation of segments defined from the mapped surface-rupture traces, with new segment activations shown as gray bars. The panels show fault-referenced and directionally organized activation rather than diffuse spatial expansion.

Figure S8 : Fault system seismicity rate and Omori-Utsu diagnostics. a–c , Fault system scale comparison between the two conjugate fault systems associated with the

M_{w}

6.4 and

M_{w}

7.1 ruptures. Panel a maps event assignments to the two fault systems, panel b shows the timing of

M\geq 3

events with symbol size scaled by magnitude, and panel c compares hourly and cumulative seismicity rates. d–f , North and south partitioning within the

M_{w}

7.1 fault system. Panel d maps the southern and northern sections, panel e shows the timing of

M\geq 3

events within these sections, and panel f compares their cumulative and hourly

M\geq 3

counts. g–i , Omori-Utsu diagnostics for the

M_{w}

7.1 fault system. Panel g shows a power-law rate fit for the full

M_{w}

7.1 fault system, with

p=0.78

and a 95% confidence interval of 0.67–0.96. Panel h shows cumulative window decay parameter estimates for the full, southern and northern sections. Panel i shows cumulative residuals between observed and expected

M\geq 3

counts for the southern and northern subsets. These diagnostics indicate heterogeneous seismicity rate evolution within the

M_{w}

7.1 fault system rather than a single spatially uniform aftershock decay.

Figure S9 :

b

-value stress proxy diagnostics for the Ridgecrest

M_{w}

6.4 to

M_{w}

7.1 interval. a , Definition of 5 km comparison regions centered on the

M_{w}

6.4 and

M_{w}

7.1 reference locations. b , Fixed window

b

-value estimates for the two comparison regions before and after the

M

5.4 separator and over the full interval between the

M_{w}

6.4 and

M_{w}

7.1 earthquakes. The

M_{w}

7.1 estimate before the separator is exploratory because of limited event support, whereas the estimate after the separator remains lower than the

M_{w}

6.4 comparison region (

\Delta b=-0.285

) and below the background level over the previous two years. c , Sliding window dynamic

M_{c}

b

-value estimates from 100-event windows. d , Spatial

b

-value field after

M

5.4, estimated with a 5 km radius and fixed

M_{c}=1.5

. e , Event support for the spatial

b

-value estimates, shown as the number of events above

M_{c}

; valid nodes require

n\geq 30

. f , Along-strike

b

-value profile on the

M_{w}

7.1 fault system, with an inset showing radius sensitivity of the comparison region estimates. The results document a relative low

b

stress proxy contrast in the

M_{w}

7.1 comparison region after

M

5.4, while treating the

b

-value estimate as complementary evidence.

TRACE: A Multi-Agent System for Autonomous Physical Reasoning for Seismology

Abstract

Figures & tables

Explore similar work

TRACE: Trajectory Reasoning through Adaptive Cross-Step Evidence Aggregation for LLM Agents

VeriTrace: Evolving Mental Models for Deep Research Agents

OpenDiscoveryTrace: Process Traces for Evaluating AI Scientist Workflows