The GOES-R series (GOES-16 to GOES-19) represents the most recent generation of NOAA's geostationary satellites, designed for continuous weather and space weather monitoring. GOES-16 (launched in November 2016, earlier GOES-East) is located at 75.2°W. GOES-17 (launched in March 2018) served as GOES-West at 137.2° W from February 2019 until January 2023 (Loto'aniu et al., 2023), after which it was relocated to on-orbit storage at 105° W.¹ GOES-18 (launched in March 2022) transitioned to 137.0° W in mid-2022 and fully assumed operational GOES-West status by early January 2023 (Loto'aniu et al., 2023). Each spacecraft carries a pair of fluxgate magnetometers mounted on a boom, with the inboard (IB) sensor positioned 6.3 m and the outboard (OB) sensor 8.5 m from the spacecraft. The IB sensor is more susceptible to thermal influences from the spacecraft bus, whereas the OB sensor generally provides greater long-term stability (Loto'aniu et al., 2023).

Following its launch in March 2022, GOES-18 was positioned close to GOES-17 for 2.5 months, with a longitudinal separation of only 0.4 degree (136.8°W vs. 137.2°W). This first of its kind collocation provided a unique opportunity for direct cross-satellite calibration, as both spacecraft observed nearly identical geomagnetic conditions. These comparisons demonstrated that GOES-18, equipped with the new Goddard magnetometers (GMAG), exhibits diurnal and long-term stability within ±1 nT, compared to variations of ~2 nT on GOES-17 and >10 nT on GOES-16 (Loto'aniu et al., 2023). This improvement reflects both the change in vendor (NASA Goddard vs. MEDA for GOES-16/17) and significant engineering upgrades, including redesigned sensor and electronics units, added heaters and thermal isolation spacers, enhanced blanketing, and extensive ground and on-orbit thermal testing, all of which mitigated the thermal instabilities seen in earlier MAG instruments (Loto'aniu et al., 2023).

The arcjet thrusters aboard the GOES-R series satellites—GOES-16 to GOES-19—are periodically fired to maintain the spacecraft's geostationary orbit. Each satellite is equipped with four thrusters, numbered 13, 14, 15, and 16 (Figure 1), nominally fire in alternating pairs (13–15 or 14–16) approximately every four days for around 90 minutes per maneuver (Califf et al., 2020a), although the exact pairing may occasionally vary. During these firings, the GOES-R magnetometers detect a significant artificial disturbance in the local magnetic field, introducing deviations of up to ~20 nT, 20% of the typical geomagnetic field strength at geostationary orbit (Califf et al., 2020b). This contamination occurs due to two primary physical mechanisms: (i) the diamagnetic effect of the thruster plume, which reduces the local ambient magnetic field strength, and (ii) plasma pressure gradients within the thruster exhaust that introduce additional localized magnetic perturbations (Califf et al., 2020b,a). The disturbance is most pronounced along the thrust axis, primarily affecting the poleward (P) component of the measured magnetic field, and remains relatively stable throughout the burn period. Although rapid recovery of the magnetic field occurs once the thrusters are deactivated, small residual offsets (~1-2 nT) may persist for hours due to thermal effects on the magnetometer electronics (Califf et al., 2020a).

Similar contamination in magnetic field measurements can also be observed in GOES-17 and GOES-18 (Figure 2). During the GOES-17/18 collocation described above, both satellites measured the Earth's magnetic field under nearly identical conditions, enabling direct cross-satellite comparisons. Outside this overlap, the longitudinal separation between the satellites leads to differences in the ambient field, preventing one from being used as ground truth for the other.

During the collocation, we identified 21 GOES-17 arcjet firing events during which GOES-18 did not fire its arcjets simultaneously. This situation provided us with a great opportunity to directly compare the effects of arcjet contamination in GOES-17 measurements. To avoid ambiguity, only events that occurred during geomagnetically quiet to moderate conditions (generally Kp ≤ 4, with no major disturbances such as magnetopause crossings) were selected. Any background wave activity present during the collocation would have been observed nearly identically by both satellites (given their 0.4 degree longitudinal separation) and therefore does not introduce bias in the cross-satellite correction. All magnetic field data are analyzed in the Earth-Pointing Normal (EPN) coordinate system, where the X-axis points toward Earth's center, the Y-axis is aligned opposite the solar panel axis (approximately anti-sunward), and the Z-axis completes the right-handed system, generally pointing northward along the spacecraft's orbital normal. This coordinate system is spacecraft-fixed and provides a stable reference frame for interpreting arcjet-related disturbances, which are closely aligned with specific spacecraft structures and thruster directions. When the arcjets are activated, their current (Figure 2c) and voltage (Figure 2d) levels increase and remain elevated while firing. Simultaneously, there is a clear reduction in the magnetic field measurements in the E and P components of the EPN coordinate frame, which persists for the full firing duration (Figures 2a, b).

Given the significance of GOES magnetometer data for scientific research, mitigation of arcjet-induced contamination is crucial. A correction algorithm was developed for both the OB and IB magnetometers on each GOES-R series spacecraft, based on an observed linear relationship between the arcjet disturbance and the ambient magnetic field (Califf et al., 2020a). While both sensors are corrected independently, OB data are generally prioritized due to its improved thermal and bias stability. This least-squares regression correction was initially found to reduce residual errors to below 1.5 nT during steady-state firings in GOES-16 (Califf et al., 2020a). However, in cases involving non-nominal firing configurations, such as single-thruster firings or evolving thrust levels, residuals from the matrix correction can be substantially larger, particularly on GOES-17 (Califf et al., 2020a). These abnormal cases are relatively uncommon compared to the standard paired-thruster maneuvers but illustrate the need for more robust correction techniques. Our study focused on well-defined paired thruster firings during the GOES-17/18 collocation; therefore, the performance of XGBoost in these rarer abnormal cases remains to be evaluated. Importantly, the existing correction matrix does not account for the short-lived transient effects at the start and end of arcjet burns, which are flagged as invalid in operational data. In designing our approach, we developed the XGBoost model to better capture these transient shoulders, providing a framework that can in principle address such short-duration features.

XGBoost (Extreme Gradient Boosting) is an advanced, scalable tree boosting system designed for efficiency, accuracy, and performance in large-scale machine learning applications (Chen and Guestrin, 2016). It is an implementation of gradient boosted decision trees (Friedman, 2001) that introduces several optimizations to enhance speed and scalability. XGBoost iteratively constructs an ensemble of decision trees, where each new tree corrects errors made by previous trees, optimizing an objective function through gradient descent (Chen and Guestrin, 2016). The system incorporates a regularized learning objective to control model complexity and mitigate overfitting, ensuring generalization to unseen data.

To develop XGBoost models for each EPN coordinate frame component, we used data from the GOES-17/18 collocation period. We selected 21 GOES-17 arcjet firing events during which GOES-18 had no simultaneous firings, ensuring a clean reference for training and validation. GOES-18 magnetic field measurements were used as the ground truth, and data from both IB and OB magnetometers were combined to increase the training volume and enhance generalizability.

Before training the XGBoost models, we preprocessed the data to make sure that the GOES-18 data, which is used as the ground truth, and the GOES-17 data, which will be corrected, overlap without any longer-term effects due to the slightly higher thermal sensitivity of the GOES-17 magnetometers (Figure 3). To achieve this objective, we first subtracted the GOES-18 measurements in each component of the EPN coordinate frame from the GOES-17 measurements during the periods when there are no arcjet firings (Figures 3a, d) for each of the 21 days from the IB and OB magnetic field data. Subsequently, we fitted a piecewise linear regression function to each difference to determine the longer-term difference between the two satellites (Figures 3b, e). We then subtracted this longer-term trend in magnetic field measurements from the GOES-18 data for each component in the EPN frame (Figures 3c, f) to create “adjusted GOES-18” values that effectively represent what GOES-17 would have measured in the absence of arcjet contamination. These adjusted values are no longer the raw GOES-18 measurements but serve as a proxy ground truth for training and validation.

Training, validation, and test sets were created from 21 GOES-17 arcjet events during the collocation period, using adjusted GOES-18 data as ground truth and GOES-17 current and voltage as inputs. We restricted the inputs to spacecraft-intrinsic parameters (current and voltage), as these directly drive the contamination mechanism and keep the correction algorithm self-contained. External variables such as the ambient magnetic field or interplanetary conditions were not included, because GOES satellites are normally located at different longitudes and the 2.5-month collocation period did not provide enough variability in interplanetary and ambient magnetic fields to train a generalized model that incorporates broader space weather effects.

Each “dataset” refers to a time window centered on a single arcjet firing, typically spanning the duration of the maneuver with a buffer before and after. This process yielded 30 arcjet events across 15 days for training, 4 events from 2 days for validation, and 4 events from 4 days for testing. Only IB magnetometer data were used for testing, as IB and OB arcjet responses are highly similar. To ensure test-day independence, all OB data from the test days were excluded from the training and validation sets.

Hyperparameter optimization was explored using a Bayesian search over a range of values (Table 1). However, optimization did not yield significant improvements compared to the default values, largely due to the limited size and variability of the training dataset. Therefore, we retained the default values in all further analyses.

In addition to the performance of the model, it is important to consider uncertainties at the instrument and platform levels that may influence the corrected results.

First, daily and seasonal changes in the thermal environment in geostationary orbit can cause diurnal structure and long-term drifts on the raw measurements, especially for the earlier R-series MAG sensors. Intersatellite analyses show that GOES-16 OB data exhibit artificial diurnal variations of order ±3 nT (1σ ≈ ±1.5 nT), whereas GOES-17 OB exhibits minimal daily variation and long-term stability within ~±2 nT (Rich et al., 2024). By contrast, the GOES-18 GMAG demonstrates much improved thermal stability: on-orbit OB-IB differences remain ≲±0.2 nT under diurnal cycling, and overall accuracy meets the NOAA ±1 nT requirement (excluding arcjet periods) (Loto'aniu et al., 2023).

Second, zero-level (offset) calibration and its refinement can introduce step-like biases if not updated. For GOES-17, yaw-flip maneuvers (180° rotations) were used to refine the OB zero offsets (e.g., −0.25 nT in P and +1.82 nT in N) and reduce long-term OB differences with other GOES spacecraft (Rich et al., 2024). Related analyses also indicate component-dependent biases on GOES-15 that manifest as pre/post-yaw-flip shifts (Rich et al., 2024). During the GOES-17/18 collocation, IB-OB statistics further show a mean offset in the GOES-17 N component (IB−OB ≈ −2.1 ± 0.62 nT), while GMAG (GOES-18) IB-OB means are near zero (Loto'aniu et al., 2023). Additionally, GOES-17 experienced a 2021 safehold event after which a small residual bias shift persisted (Loto'aniu et al., 2023).

Third, sensor placement and thermal coupling lead to IB/OB differences. Prior work showed the GOES-16 IB sensor is more thermally susceptible than OB, producing time-varying offsets of several nT tied to eclipse seasons and diurnal heating; ML corrections reduced these variations from ~3-5 nT to ~0-2 nT (E component) but some residual, MLT-dependent offsets remained (Inceoglu and Loto'aniu, 2021). These findings motivate our use of GOES-18 as the reference and our emphasis on spacecraft-intrinsic inputs for the correction.

In the context of our study, these factors imply that (i) residual uncertainty in the corrected GOES-17 series is bounded below by the stability/accuracy of the GOES-18 GMAG (order ~1 nT) and by any residual, component-dependent GOES-17 biases during the collocation; and (ii) IB and OB arcjet responses are highly similar in our data, but long-term IB/OB thermal susceptibilities differ. Practically, this is why we (a) trained against adjusted GOES-18 (collocation-aligned) values, (b) withheld OB data from test days, and (c) evaluate performance by component and day. We recommend users treat start/end-of-burn transients and days with known configuration changes or post-anomaly bias shifts with additional caution, and we view uncertainty quantification (e.g., confidence intervals on corrections or QC flags) as a natural extension for future operational use.

It should be noted that the XGBoost correction was designed to address plume-related arcjet contamination during firing periods and does not explicitly correct for the small residual thermal offsets (~1-2 nT) that can persist for hours after thruster shutdown (Califf et al., 2020a). These long-lived effects were removed during preprocessing and remain a separate calibration challenge.