Analysis of Induced Seismicity Characteristics and Mitigation Strategies in the Development of Hot Dry Rock Geothermal Resources: A Review

Abstract

Induced seismicity is a recognized challenge in hot dry rock (HDR) geothermal development. Based on a systematic review of previous studies on induced seismicity mechanisms, characteristics, risk assessment, and mitigation measures, we compiled publicly available data from 13 HDR projects worldwide. Our statistical analysis shows that felt earthquakes occurred in 62% of the projects, and a post-injection “tailing effect” was observed in 54% of the projects. The spatial influence range of induced seismicity is typically within 2 km of the injection well, and the duration of the “Kaiser effect” varies from months to years depending on local conditions. A cross-site comparison of injection parameters suggests that the maximum wellhead pressure may be a more useful indicator than injected volume for estimating the largest possible earthquake magnitude, especially when comparing different tectonic settings. Furthermore, we examine the applicability and limitations of b-value trends, seismogenic indices, and existing maximum magnitude prediction models in seismic risk assessment. Dynamic adjustment of injection parameters based on real-time risk indicators, combined with safer injection schemes, may represent an important research direction for improving the conventional traffic light system. These findings provide a data-driven basis for site-specific safety management of HDR development.

1. Introduction

Hot dry rock (HDR) denotes high-temperature geological formations located at depths of several kilometers that contain little to no native fluid. Owing to its vast resource base and renewable nature, HDR constitutes a low-carbon geothermal energy option with substantial long-term potential [1]. Commercial utilization relies on Enhanced Geothermal Systems (EGS), in which hydraulic stimulation creates fracture networks within low-permeability reservoirs to establish closed-loop circulation. A working fluid is injected through injection wells; it extracts heat as it migrates through the stimulated volume and is subsequently produced via production wells for power generation (Figure 1). The earliest HDR field trials began in 1973 at Fenton Hill, United States, followed by experimental campaigns in the United Kingdom, France, Japan, Germany, and Australia. Among these, the Soultz-sous-Forêts project (France) achieved the first sustained commercial-scale electricity generation [2,3]. In 2019, the U.S. Department of Energy launched the Frontier Observatory for Research in Geothermal Energy (FORGE) initiative to reinvigorate fundamental and applied HDR research. In the same year, China initiated field-scale trials in Gonghe (Qinghai Province), Matouying (Tangshan), and Xinghua (Jiangsu Province). Collectively, these efforts indicate that HDR geothermal resources are poised to serve as a strategic complement to existing energy portfolios and may substantially reshape future energy supply configurations.

Figure 1. Schematic diagram of hot dry rock development [4].

The expanding deployment of hydraulic stimulation in HDR developments, coupled with improved seismic monitoring capabilities, has revealed that injection-induced seismicity constitutes a non-negligible safety concern. Multiple HDR projects have already been suspended or terminated because of the unacceptable seismic risk. Induced events are thus emerging as a primary constraint on the scalable exploitation of HDR resources.

Although previous review articles have systematically summarized the mechanisms and characteristics of induced seismicity [5], prediction methods [6], criteria for distinguishing induced from natural earthquakes, and risk management frameworks [7,8] in enhanced geothermal systems (including both hot dry rock and hydrothermal systems), cross-site comparative analyses focusing specifically on the induced seismicity characteristics of HDR developments remain limited. In particular, the relationship between injection parameters and induced seismicity has not been well synthesized, making it difficult to directly transfer existing knowledge to seismic risk assessment for new projects. Moreover, induced seismicity data from HDR projects developed over the past decade—such as Pohang in South Korea, Gonghe and Tangshan in China, and FORGE in the USA—have not yet been integrated into a unified comparative framework.

To address this gap, we compiled publicly available data from 13 HDR projects worldwide, summarized the induced seismicity characteristics of major development sites, and provided basic parameters such as the spatial influence range of induced seismicity and the duration of the “Kaiser effect”. Cross-site comparison of injection parameters indicates that the maximum wellhead pressure may be a more informative indicator than injected volume when evaluating the maximum possible magnitude. This relationship is particularly pronounced in reservoirs with high pressure and low injectivity. We also discuss the applicability and limitations of major seismic risk indicators (including b-value, seismogenic index, and seismic release efficiency) as well as existing maximum-magnitude prediction models. On this basis, we further explore the potential value of real-time adjustment of injection parameters—based on real-time risk indicators—for improving conventional traffic light systems. The above analyses aim to provide a data-driven reference for site-specific safety management of HDR development.

2. Seismic Characteristics Induced by HDR Development

2.1. Data Compilation and Analysis Methodology

All data used in this study came from published peer-reviewed literature and publicly available project reports. To systematically compile induced seismicity data from global HDR projects, we searched Web of Science, Scopus, Google Scholar, CNKI, and relevant project reports (e.g., Gonghe), covering the period from the earliest available report of each project to December 2024. Keywords included “hot dry rock,” “enhanced geothermal system,” “induced seismicity,” “hydraulic fracturing,” and their combinations. Given the limited number of publicly documented HDR field sites, we additionally employed a site-driven supplementary search strategy: we first identified all known HDR/EGS test sites, then performed targeted searches using their formal and informal site names, project names, and associated geological formation names in the databases and grey literature. We also conducted backward and forward citation tracking of key papers. This two-pronged strategy ensured near-complete coverage of all publicly documented HDR projects with induced seismicity records.

For quality assessment, priority was given to peer-reviewed journal papers and authoritative project reports. When multiple sources existed for the same project (e.g., Soultz, Basel), we used the most recent and complete synthesis report, cross-validating key parameters such as maximum magnitude and injection pressure. Injection parameters (injection pressure, injection rate, cumulative injected volume), induced event magnitudes (M_L or M_w), b-values, and spatial extents of seismic clouds were extracted directly from the original literature. No original seismic waveforms or raw injection records were reprocessed or recalculated.

The site selection criteria were as follows: (1) the project was explicitly reported as an HDR development; (2) at least three of the following parameters were reported in the open literature: maximum induced magnitude, wellhead pressure, injection flow rate, or cumulative injected volume; and (3) basic statistical information on induced seismic events (e.g., magnitude distribution, b-value, seismic cloud extent) was provided. Based on these criteria, 13 sites were selected. To ensure consistency across sites, priority was given to the moment magnitude (M_w) as reported in the original literature. For events reporting only local magnitude (M_L)—most of these reports come from France and Switzerland—we uniformly applied the conversion formula of Grünthal et al. (2009) [9], which was derived from regression on central and northern European seismic data: M_w = 0.0376 × M_L² + 0.646 × M_L + 0.53. Data items marked with “-” in Table 1 mean that the parameter was not reported in the original source. In correlation and regression analyses, only complete pairs of data were retained, and missing values were excluded from statistical tests. For all linear regressions, 95% confidence intervals were calculated using the bootstrap method with 1000 resampling iterations. Because the sample size is small and the data may not follow a normal distribution, we used Spearman’s rank correlation coefficient to evaluate the monotonic relationship between injection parameters and maximum magnitude. For the bimodal trend between injection pressure and maximum magnitude, Spearman rank correlation analyses were performed separately for the low-pressure group (<30 MPa) and the high-pressure group (≥30 MPa) to assess the correlation between injection pressure and maximum magnitude within each group. The goodness of fit of each regression model is reported as R² and p-values. All statistical analyses were performed using SPSS 22.0 software.

2.2. Event Magnitudes

A systematic compilation of publicly available data from thirteen major HDR stimulation campaigns worldwide indicates that eight sites (62%) generated perceptible events with local magnitude M ≥ 2. At four of these sites (31%), the largest events exceeded M3. The highest-magnitude event documented to date is the M_w 5.5 Pohang earthquake (Republic of Korea); however, the extent to which this event can be attributed to HDR operations remains contested.

For comparison, hydraulic fracturing in conventional hydrocarbon reservoirs and stimulation of hydrothermal geothermal systems typically produce maximum moment magnitudes below M_w 1 [10]. The propensity of HDR reservoirs—hosted predominantly in deep, competent crystalline rocks—to generate comparatively larger events is consistent with the higher differential stresses, lower fracture compliance, and limited poroelastic relaxation characteristic of such lithologies.

2.3. Spatial Extent of Influence

The distribution of microseismicity delineates the zone affected by hydraulic stimulation and provides a quantitative basis for site-specific safety assessments and fault setback distances. Numerical modeling by Westwood et al. [11] suggests that maintaining a horizontal separation of at least 433 m between the injection interval and a nearby fault minimizes the likelihood of perceptible induced seismicity. Using microseismic catalogues, Davies et al. [12,13] propose a vertical separation of 600 m between stimulated reservoirs and overlying aquifers to prevent fluid migration and contamination. Wilson et al. [14] analyzed 108 fluid-injection operations and reported a maximum horizontal detection distance of 895 m for microseismic events. The present compilation of HDR field data (Table 1) indicates that the horizontal extent of microseismic clouds ranges from 0.1 to 2 km, while the vertical influence spans from 0.2 to 1.5 km. These dimensions exceed those typically observed in conventional hydraulic fracturing of sedimentary formations, reflecting the deeper, stiffer crystalline host rocks and the larger injected volumes characteristic of HDR projects. Moreover, poroelastic stress transfer and aseismic slip can extend the seismogenic footprint beyond the hydraulically stimulated reservoir volume [15,16], underscoring the need to consider these processes when defining exclusion zones.

Table 1. Statistical table of earthquakes induced by major hot dry rock development sites.

#	Site	Well	Time	Depth (km)	Inj. Mode	Stress	Vtot	Qmax	Pmax	Mmax-in	Mmax-p	Extent	Mc/b	Ref.
1	Basel	Basel1	2009	4.7–5	a	SS	11.5	55	29.6	2.6 ML	3.4 ML	1 × (<0.1)× 1	0.9/1.15	[17]
2	Pohang	PX1/PX2	2016– 2017	4.2/4.3	a	SS	12.8	46.8	89.2	2.0 Mw	5.5 Mw	PX1 (0.5 × 0.8 × 0.2), PX2 (1 × 0.2 × 0.5)	1.1/0.7	[18]
3	Paralana	Paralana2	2011	2.7–3.6	a	TF	3.1	27	62	2.5 Mw	/	-	−0.3/0.94	[19]
4	Cooper	Habanero#1	2003	4.3	a	TF	20	40	70	3.7 Mw	/	2 × 1.5 × 0.2	−0.9/0.77	[20]
5	Helsinki ST1	OTN-3	2018	5.5–6.1	a	SS-TF	18.2	13.3	90	1.9 Mw	/	<0.6	−0.6/1.35	[21]
6	Rosemanowes	RH12-15	1984–1992	2.7	a/b	SS	100	200	-	2 ML	-	-	-	[22]
7	Ogachi	OGC-1	1991	1.1	a	-	10.1	11	20	2 ML	/	1 × 0.4 × 0.5		[23]
8	Hijiori	HDR-1	1992	2.2	a	NF	2.11	66.6	25.5	0.3 ML	/	0.7 × 0.15 × 0.4		[24]
9	Gonghe	GH01	2020	3.6	a	SS-TF	26	52	59	1.9 ML	2.28 ML	1.1 × 0.8 × 0.8		[3]
10	Tangshan MTY	M1	2022	4.2–4.5	a	SS	87.2	49	27	-	1.9 ML	-		[25]
11	Fenton Hill	EE-2	1983	3.5	a	SS-NF	21.3	-	50	1.3 ML	/	0.95 × 0.15 × 0.8		[26]
12	FORGE Utah	16A	2022	2.5–2.6	a		1.60	140	49.8	-	0.5 M	-	-	[27]
13	FORGE Utah-1	16A/16B	2023	-	b		2.96	20	32.6	-	0.45 M	-	-
14	Soultz93	GPK1	1993	3.4–3.5	a	SS-NF	25.3	36	11.4	1.7 ML	1.9 ML	1.2 × 0.5 × 1.5		[5]
					a		19.3	50	10	1.6 ML	/	-	−1.8/1.1
15	Soultz96	GPK2	1995–1996	3.2–3.9	a	SS-NF	28	78	13	-	-	0.6	-	[28]
16	Soultz00	GPK2	2000	5	a	SS-NF	22.7	90	14.5	2.5 ML	/	-	0.1/0.94	[29]
					a	SS-NF	45	35	-	-	2.6 ML	1.5 × 0.6 × 1.2
17	Soultz03	GPK3	2003		a	SS-NF	34	50	17.9	-	-	2 × 0.7 × 1.5	-	[30]
		GPK2	2003		a	SS-NF	34	20	7.9	2.5 ML	2.9 ML		0/1.24
18	Soultz05	GPK4	2004	5.1	a	SS-NF	9	44	17	-	-	-	-	[31]
			2005		a	SS-NF	12.5	45	18	2.7 ML	/	-
19	Soultz-1	GPK1/GPK2/GPK3	2005	4.5–5	b	SS-NF	-	15	-	2.3 ML	/	-
20	Soultz-2	GPK2/GPK3	2008		b	SS-NF	-	25	-	1.4 ML	/	-

Inj. mode: (a) reservoir stimulation, (b) circulation; Stress: SS, strike-slip stress regime; NF, normal-faulting stress regime; TF, thrust-faulting stress regime; V_tot: total volume of injected fluid (×10³ m³); Q_max: maximum injection rate (L/s); P_max: maximum well head pressure (Mpa); M_max-in: Maximum magnitude during injection; M_max-p: Post-shut-in maximum magnitude; Extent: Spatial extent of the seismic cloud (length × width × height) (km); M_c: completeness magnitude, b: seismic b-value; -: data not available; /: no larger event recorded after pump shut-in; M_L: local (Richter) magnitude; M_w: moment magnitude.

Table 1. Statistical table of earthquakes induced by major hot dry rock development sites.

#	Site	Well	Time	Depth (km)	Inj. Mode	Stress	Vtot	Qmax	Pmax	Mmax-in	Mmax-p	Extent	Mc/b	Ref.
1	Basel	Basel1	2009	4.7–5	a	SS	11.5	55	29.6	2.6 ML	3.4 ML	1 × (<0.1)× 1	0.9/1.15	[17]
2	Pohang	PX1/PX2	2016– 2017	4.2/4.3	a	SS	12.8	46.8	89.2	2.0 Mw	5.5 Mw	PX1 (0.5 × 0.8 × 0.2), PX2 (1 × 0.2 × 0.5)	1.1/0.7	[18]
3	Paralana	Paralana2	2011	2.7–3.6	a	TF	3.1	27	62	2.5 Mw	/	-	−0.3/0.94	[19]
4	Cooper	Habanero#1	2003	4.3	a	TF	20	40	70	3.7 Mw	/	2 × 1.5 × 0.2	−0.9/0.77	[20]
5	Helsinki ST1	OTN-3	2018	5.5–6.1	a	SS-TF	18.2	13.3	90	1.9 Mw	/	<0.6	−0.6/1.35	[21]
6	Rosemanowes	RH12-15	1984–1992	2.7	a/b	SS	100	200	-	2 ML	-	-	-	[22]
7	Ogachi	OGC-1	1991	1.1	a	-	10.1	11	20	2 ML	/	1 × 0.4 × 0.5		[23]
8	Hijiori	HDR-1	1992	2.2	a	NF	2.11	66.6	25.5	0.3 ML	/	0.7 × 0.15 × 0.4		[24]
9	Gonghe	GH01	2020	3.6	a	SS-TF	26	52	59	1.9 ML	2.28 ML	1.1 × 0.8 × 0.8		[3]
10	Tangshan MTY	M1	2022	4.2–4.5	a	SS	87.2	49	27	-	1.9 ML	-		[25]
11	Fenton Hill	EE-2	1983	3.5	a	SS-NF	21.3	-	50	1.3 ML	/	0.95 × 0.15 × 0.8		[26]
12	FORGE Utah	16A	2022	2.5–2.6	a		1.60	140	49.8	-	0.5 M	-	-	[27]
13	FORGE Utah-1	16A/16B	2023	-	b		2.96	20	32.6	-	0.45 M	-	-
14	Soultz93	GPK1	1993	3.4–3.5	a	SS-NF	25.3	36	11.4	1.7 ML	1.9 ML	1.2 × 0.5 × 1.5		[5]
					a		19.3	50	10	1.6 ML	/	-	−1.8/1.1
15	Soultz96	GPK2	1995–1996	3.2–3.9	a	SS-NF	28	78	13	-	-	0.6	-	[28]
16	Soultz00	GPK2	2000	5	a	SS-NF	22.7	90	14.5	2.5 ML	/	-	0.1/0.94	[29]
					a	SS-NF	45	35	-	-	2.6 ML	1.5 × 0.6 × 1.2
17	Soultz03	GPK3	2003		a	SS-NF	34	50	17.9	-	-	2 × 0.7 × 1.5	-	[30]
		GPK2	2003		a	SS-NF	34	20	7.9	2.5 ML	2.9 ML		0/1.24
18	Soultz05	GPK4	2004	5.1	a	SS-NF	9	44	17	-	-	-	-	[31]
			2005		a	SS-NF	12.5	45	18	2.7 ML	/	-
19	Soultz-1	GPK1/GPK2/GPK3	2005	4.5–5	b	SS-NF	-	15	-	2.3 ML	/	-
20	Soultz-2	GPK2/GPK3	2008		b	SS-NF	-	25	-	1.4 ML	/	-

Inj. mode: (a) reservoir stimulation, (b) circulation; Stress: SS, strike-slip stress regime; NF, normal-faulting stress regime; TF, thrust-faulting stress regime; V_tot: total volume of injected fluid (×10³ m³); Q_max: maximum injection rate (L/s); P_max: maximum well head pressure (Mpa); M_max-in: Maximum magnitude during injection; M_max-p: Post-shut-in maximum magnitude; Extent: Spatial extent of the seismic cloud (length × width × height) (km); M_c: completeness magnitude, b: seismic b-value; -: data not available; /: no larger event recorded after pump shut-in; M_L: local (Richter) magnitude; M_w: moment magnitude.

2.4. Post-Injection Seismicity

The term “tailing effect” denotes the continued occurrence of seismic events after the cessation of fluid injection. These events typically nucleate at the periphery of the previously active microseismic cloud and are frequently characterized by magnitudes larger than those observed during active injection [17,18,19]. Among the thirteen HDR stimulation campaigns compiled in Table 1, seven projects (54%) exhibit unequivocal evidence of post-shut-in events exceeding the maximum magnitude recorded during injection. The lag time between shut-in and the largest post-injection event varies markedly, reflecting site-specific geological conditions and injection parameters. For example, the Denver triggered-earthquake experiment exhibited a lag of ~10 days [32], whereas seismicity at the Pohang PX-2 well (Republic of Korea) persisted for approximately 6 months. Nevertheless, the largest post-injection events in most datasets occur within three months after shut-in. The Basel EGS site registered its maximum magnitude (M_w 3.4) on the day of shut-in and experienced three additional M_w ≥ 3 events within the subsequent 3 months. Similarly, the M_w 5.5 Pohang earthquake occurred 49 days after the injection ceased.

2.5. “Kaiser Effect”

The “Kaiser effect” refers to the phenomenon that, after unloading and subsequent reloading of rocks or materials, acoustic emissions occur only when the stress exceeds the previously experienced maximum stress [33]. In hot dry rock (HDR) development, the same “Kaiser effect” is observed in induced seismicity, where seismic activity is reactivated only when the wellhead injection pressure exceeds the peak value of the previous injection [5,32,34]. At the Paralana (Australia) site, the main stimulation phase failed to reactivate seismicity until the wellhead pressure surpassed the peak value recorded during an earlier trial conducted six months prior, implying a stress-field recovery time exceeding half a year. Similarly, reinjection experiments at the Copper Basin site, separated by a two-year interval, displayed a clear Kaiser signature. Evans et al. [31] compared microseismicity induced by successive stimulations at Soultz and concluded that 4 months were insufficient for either stress-field restoration, or fracture closure, whereas a 3-year hiatus was adequate. Two injection tests performed six years apart at the KTB deep borehole further demonstrated that the majority of post-reinjection events nucleated beyond the footprint of the prior seismic cloud, suggesting that the stress field had not fully recovered even after six years. Stress-field recovery is governed by the reservoir stress state, geological architecture, injected energy, and treatment volume; consequently, recovery times vary among sites. Nevertheless, the observed Kaiser intervals across multiple HDR projects indicate that even modest injection treatments do not ensure complete stress-field restoration within six months.

3. Mechanical Mechanisms of Injection-Induced Seismicity

Fundamental insights into injection-induced seismicity originate from the fluid-injection experiments conducted in Colorado during the 1960s–1970s, where elevated pore pressure was identified as the primary driver reducing effective normal stress on pre-existing faults [35]. Subsequent investigations have expanded the understanding to encompass poro-elastic stressing, thermal stresses, earthquake-earthquake interactions, fault strength heterogeneity, and aseismic slip [35,36]. Keranen et al. [37] proposed a tripartite classification of induced-seismicity modes (Figure 2): (1) Pore–pressure increase with a consequent reduction in effective normal stress; (2) Combined influence of pore-pressure rise and poro-elastic stress perturbation; (3) Poro-elastic stress perturbation in the absence of significant pore-pressure change. Across a broad range of field observations and numerical studies, pore-pressure diffusion remains the dominant mechanism during fluid injection [37,38,39,40,41,42] whereas poro-elastic stress changes are primarily associated with fluid extraction scenarios [43,44].

Figure 2. Schematic diagram of earthquakes induced by the development of hot dry rock and the Mohr’s circle diagrams of three induced seismic mechanical mechanisms [37,45]. ① Pore-pressure increase with consequent a reduction in effective normal stress. ② Combined influence of pore-pressure rise and poro-elastic stress perturbation. ③ Poro-elastic stress perturbation in the absence of significant pore-pressure change.

The physical processes governing seismicity after injection cessation remain only partially resolved. Numerical experiments by Baisch et al. [46] demonstrate that minimal residual pore-pressure diffusion can be sufficient to drive large regions to criticality once the well is shut in. Analytical solutions provided by Segall and Lu [47] attribute post-injection events to continuing poro-elastic stress perturbations that persist after the pressure transient has decayed. Using coupled thermo-poro-elastic simulations, Simone et al. [48] show that the superposition of thermal contraction and shear-stress transfer along existing fractures can sustain, or even enhance, seismic activity after shut-in. Field evidence and complementary modelling further suggest that previously opened fractures and faults begin to re-close as fluid pressure declines. These closing surfaces act as hydraulically impermeable barriers, redirecting residual fluids toward previously unstimulated zones at the periphery of the reservoir [48,49,50]. The resulting pressure build-up along these distal structures lowers the effective normal stress. It can trigger failure, thereby accounting for the systematic observation that post-injection events predominantly nucleate outside the footprint of the earlier seismic cloud.

4. Discrimination Between Induced and Natural Seismicity

Reliable discrimination between induced and tectonic earthquakes remains a prerequisite for seismic hazard assessment and operational control of HDR projects. Yet, no single seismological criterion has been established that can unambiguously separate the two populations [37]. At low magnitudes (M < 0), hydraulic-fracturing events may exhibit an increased tensile component in their focal mechanisms [51,52]; however, moderate-to-large events display source mechanisms, spectral content, stress-drop values, and ground-motion characteristics that are statistically indistinguishable from those of shallow crustal earthquakes [42]. Consequently, waveform-based or source-parameter-based identification alone is insufficient for felt events. Current practice therefore relies on spatiotemporal association tests and departures from the regional seismic baseline [10,53,54]. Zhang et al. [55] further emphasize that statistically significant correlations between seismicity metrics and injection parameters (e.g., injection rate, cumulative volume, wellhead pressure) should be incorporated as an additional discriminatory criterion beyond simple spatiotemporal proximity.

5. Statistical Analysis of Induced Seismicity Versus Injection Parameters Across Field Sites

5.1. Relationship with Injected Volume

Inspection of Figure 3a reveals no universal scaling between the maximum induced magnitude and injected volume when data from different sites are pooled. However, within a single site, the largest event per stimulation stage exhibits an approximately linear trend on a semi-logarithmic plot of magnitude versus injected volume [56]. Earlier studies reported that cumulative seismic moment increases linearly with total injected volume [38], and Shapiro et al. [57] demonstrated that the number of induced events scales proportionally with injected volume. Conversely, Van der Elst et al. [58] used observational residuals, the absence of a robust correlation between cumulative moment and volume, and inconsistencies with physical scaling arguments to refute the proportionality between cumulative moment and injected volume, while reaffirming the proportionality between event count and volume.

Figure 3. Relationship between maximum magnitude and injection parameters. The numbers in the figure correspond to the project serial numbers in Table 1. (a) Injected volume vs. maximum magnitude; (b) maximum injection rate vs. maximum magnitude; (c) maximum wellhead pressure vs. maximum magnitude; (d) injectivity vs. maximum magnitude, where injectivity is roughly estimated as the ratio of maximum injection rate to maximum wellhead pressure.

5.2. Relationship to Injection Flow Rate

Field data (Figure 3b) show that HDR stimulations typically employ flow rates below 50 L/s, yet no clear dependence of the maximum magnitude on peak flow rate is apparent. Alghannam and Juanes (2020) [59] experimentally demonstrated that for a fixed cumulative volume, abrupt increases in injection rate elevate seismic hazard, whereas gradual ramping reduces it. These findings indicate that rapid rate changes generate sharp pore-pressure transients, which in turn increase the probability of triggering larger events.

Cross-site analysis (Figure 3b) indicates that neither the magnitude nor the rate of induced seismicity exhibits a statistically significant, direct correlation with peak injection flow rate. Most HDR stimulations are conducted at flow rates below 50 L/s, and the largest events do not cluster at higher rates. Laboratory and field experiments by Alghannam and Juanes (2020) [59]; however, show that for a fixed cumulative volume, abrupt increases in flow rate markedly elevate seismic hazard, whereas gradual ramping attenuates it. The authors attribute the heightened risk to rapid pore-pressure transients generated by sudden rate changes, which can drive fractures or faults to critical stress conditions more efficiently than monotonic or stepwise injection schedules.

5.3. Relationship to Injection Pressure

Induced seismicity frequency exhibits a strong positive correlation with injection pressure, a pattern consistently observed across both HDR developments and controlled injection experiments [29,60]. The underlying mechanism is straightforward: once the target fracture network is hydraulically connected to the wellbore, the elevated wellhead pressure increases pore pressure, reduces effective normal stress, and thereby promotes slip on pre-existing fractures or faults. Inspection of the relationship between maximum magnitude and peak injection pressure reveals a clear bimodal distribution (Figure 3c).

Group 1 (low-pressure group) includes sites such as the multi-stage stimulation wells at Soultz (France) and the Basel project (Switzerland). These reservoirs are located in the Rhine Graben and its adjacent extensional tectonic zones, where the stress regime is predominantly strike-slip to normal-faulting, the tectonic setting is relatively consistent, and the injectivity is relatively high (>1 L/MPa·s, see Figure 3d). In this group, injection pressure shows a significant positive correlation with the maximum magnitude (Spearman ρ = 0.93, p < 0.01), with a bootstrap 95% confidence interval of [0.41, 1.00]. Notably, when the Japanese Ogachi site—which visually also belongs to this group—is included in the analysis, the correlation disappears (ρ = 0.29). This further suggests that the good correlation observed within the low-pressure group may be mainly controlled by regional tectonic homogeneity rather than by a simple pressure threshold effect.

Group 2 (high-pressure group) covers sites such as Fenton Hill and Utah (USA), Cooper Basin and Paralana (Australia), Pohang (Republic of Korea), and Gonghe (China). These sites share similar characteristics: a strike-slip to reverse-faulting stress regime, high crustal stress, and low injectivity (0.41–0.91 L/MPa·s, see Figure 3d). In this group, injection pressure exhibits an extremely strong monotonic positive correlation with the maximum magnitude (Spearman ρ = 1.00, p < 0.01), with a bootstrap 95% confidence interval of [1.00, 1.00]. Notably, the Pohang M_w 5.5 event (Republic of Korea)—although widely recognized as having reactivated a pre-existing fault and exceeding the predictions of injection-volume-based models (e.g., McGarr, 2014 [38])—still falls within the pressure-magnitude trend of this group. Another exception is the Finnish ST1 project: this site also has low injectivity and a high wellhead pressure (90 MPa), but it did not generate a high-magnitude event (maximum magnitude only about 1.9). Its main measures included real-time adjustment of injection parameters based on seismic risk analysis, such as controlling wellhead pressure and extending injection intervals. This case demonstrates that even under high-pressure, low-injectivity conditions, real-time adjustment of injection parameters can effectively suppress the growth of induced magnitudes, providing a feasible path for active risk management.

In summary, the “bimodal” distribution observed in this study is not simply a pressure threshold effect. The available field data suggest that this phenomenon may result from the combined influence of different tectonic stress regimes and injectivity. The above analysis further indicates that injection pressure may be a more robust and more valuable engineering early-warning indicator than injected volume when evaluating the maximum magnitude. At the same time, real-time adjustment of injection parameters offers a pathway toward safer hot dry rock development.

6. Seismic-Hazard Assessment of Induced Events

6.1. Conditions for Generating Perceptible Induced Earthquakes

The prevailing consensus holds that four necessary conditions must be simultaneously satisfied for fluid injection to generate a perceptible (felt) earthquake [45,61]: (i) an active stress perturbation source; (ii) a pre-existing fault at or near critical stress; (iii) a fault of sufficient geometric extent to produce a felt event; and (iv) a hydromechanical coupling pathway linking the stress perturbation source to the fault. Mechanically, once a fault intersects the stimulated reservoir volume—thereby establishing direct hydraulic communication—elevated injection pressures can destabilize faults even if their initial stress state is sub-critical. The 2011 Prague, Oklahoma, earthquake sequence exemplifies this behavior: ruptures nucleated not only along favourably oriented (critically stressed) planes but also along misoriented (non-critically stressed) structures [36]. These observations confirm that elevated pore pressures expand the range of fault orientations susceptible to failure [45].

6.2. B-Value Characteristics

The b-value, defined as the slope of the Gutenberg-Richter magnitude-frequency distribution, typically approximates 1.0 for tectonic earthquakes [62]. High-pressure fluid-injection experiments generally yield b-values significantly greater than 1.0 [63]. In contrast, induced seismicity recorded at HDR sites exhibits lower b-values, most of which cluster around or below 1.0 (Table 1). No systematic correlation is observed between the maximum induced magnitude and the b-value. Although the M_w 5.5 Pohang (Republic of Korea) and the large-magnitude sequence in the Cooper Basin (Australia) are both characterized by below-average b-values, the two datasets exhibit comparable b-values despite markedly different peak magnitudes. Consequently, the absolute b-value does not serve as a quantitative predictor of seismic hazard. Rather, because the b-value reflects the relative proportion of small to large events, a sustained decrease toward or below 1.0 can be interpreted qualitatively as an indicator of an elevated potential for larger-magnitude events.

6.3. Seismic-Energy Release Efficiency

Maxwell et al. [64] introduced the seismic release efficiency (I_E) as:

$${\mathit{I}}_{\mathit{E}}=\mathit{\Sigma }\mathit{E}\mathit{o}/{\mathit{E}}_{\mathit{H}}$$

(1)

where ΣE_O is the cumulative radiated seismic energy and E_H is the hydraulic energy. Bentz et al. [56] compared I_E across multiple EGS sites and found that most projects converge within the range from 10⁻³ to 10⁻². Projects that ultimately produced larger events—specifically the Cooper Basin (Australia) and Pohang (Republic of Korea)—experienced rapid increases in I_E. The Pohang sequence is particularly notable: I_E was initially low but escalated sharply prior to the M_w 5.5 event.

Conversely, the Finnish ST1 project maintained I_E at the lower end of the observed spectrum, successfully limiting the maximum magnitude to 2.0 [21]. Nevertheless, the Basel stimulation demonstrates that I_E alone is insufficient as a safeguard: I_E declined monotonically throughout the operation, yet the sequence culminated in an M_L 3.4 earthquake. Consequently, while constraining I_E can reduce seismic hazard, it must be embedded within a broader, multi-parameter risk-management protocol.

6.4. Seismogenic Index

Shapiro et al. [57] introduced the seismogenic index (Σ) to quantify the propensity of a given geological setting to generate injection-induced seismicity. Derived under the assumption that pore-pressure diffusion is the dominant triggering mechanism and incorporating the Gutenberg–Richter distribution, the index is independent of operational injection parameters and depends solely on tectonic characteristics:

$$\mathit{\Sigma }=\mathit{l}\mathit{o}\mathit{g}\mathit{N}\left(\mathit{t}\right)-\mathit{l}\mathit{o}\mathit{g}\mathit{Q}\mathit{c}\left(\mathit{t}\right)+\mathit{b}\mathit{M}$$

(2)

where N(t) is the cumulative number of events with a magnitude of ≥M, Qc(t) is the cumulative injected volume, and b is the b-value of the Gutenberg–Richter distribution.

HDR projects exhibit systematically higher Σ values compared with the hydrocarbon stimulations. Specifically, Cooper Basin (Australia) records Σ ≈ −1.2, and Basel (Switzerland) records Σ ≈ 0.1; both sites experienced events exceeding M_w 3. In contrast, Ogachi (Japan) displays Σ ≈ −2.6, and it generated significantly lower-magnitude seismicity. These observations indicate that higher seismogenic indices correspond to elevated seismic hazard under equivalent injection volumes.

6.5. Alignment of Seismic-Cloud Planes

Large, induced events are frequently observed to nucleate on discrete planar structures that can be resolved within high-resolution microseismic clouds. During the circulation phase at Soultz (France), the largest event (M_L 2.9) occurred on the F3 fault plane, as interpreted from the seismic cloud [65]. Similarly, the Pohang project (Republic of Korea) intersected a high-angle fault that caused severe mud losses while drilling; microseismicity was first triggered at this location, and the subsequent M_w 5.4 earthquake ruptured the same plane [18]. Because deep-seated faults—particularly steeply dipping structures in crystalline basement—are difficult to detect a prioritised, real-time, high-precision microseismic monitoring should be employed throughout HDR development to identify and continuously update the geometry of seismogenic planes. The interpretation of these seismic-cloud planes is a critical qualitative component of induced-seismicity risk assessment.

6.6. Maximum-Magnitude Forecasting

Two conceptually distinct upper bounds are invoked for the largest possible magnitude (M_max) that can be induced during HDR stimulation [45]: (1) the largest earthquake that the seismogenic volume could host under purely natural conditions, as inferred from regional seismicity and tectonic assessment [39,44] and (2) a volume-dependent bound dictated by the injected fluid volume within the stimulated reservoir [57,66]. For the second bound, statistical and reduced-physics models have been developed that couple the spatial distribution of microseismicity with simplified fracture-mechanics considerations and have shown reasonable agreement with field observations. The three most frequently applied forecasting frameworks are: the McGarr model [38]; the Van der Elst model [58] and Galis model [67]. When the three models are benchmarked against the HDR projects listed in Table 1 (Figure 4), the McGarr (fixed b = 1) and Galis (γ = 1.5 × 10⁸) formulations systematically over-predict observed M_max. The McGarr overestimation is attributable to its implicit neglect of aseismic deformation and enhanced permeability, which reduce seismic moment release per unit of volume. Although conservative, the McGarr estimate is therefore useful as an upper bound when no other constraints are available. The Van der Elst model, which depends on both the seismogenic index (Σ) and the b-value, is more sensitive to parameter choice. Its statistical premise—that Σ and b remain temporally invariant—limits applicability in heterogeneous or evolving stress environments. Nevertheless, because it leverages real-time seismicity statistics, it generally produces more accurate short-term forecasts during active injection. The Galis formulation is strictly a mechanical bound for self-arrested rupture rather than a predictive model for induced M_max. The observation that all recorded HDR-induced events fall below this bound merely indicates that ruptures to date have been self-limiting, implying technical controllability under current operational envelopes.

Figure 4. Comparison chart of maximum magnitude predictions from models versus observed maximum magnitudes at various hot dry rock development sites. The red line represents the prediction curve of the McGarr model, the black line represents the prediction curve of the Galis model, and the green curves represent the prediction curves of the Van der Elst model under different seismogenic index values.

Current prediction models are mostly based on simplified physical frameworks [6]. Developing more accurate multi-field coupling models is an important direction for future maximum magnitude forecasting. Recent studies have highlighted the key role of temperature effects on the reservoir pore pressure, fluid viscosity, and rock mechanical properties in physical prediction models of induced seismicity [6,68]. Therefore, in multi-field coupling models, more attention should be paid to temperature effects, for instance, by incorporating non-isothermal flow models [69], to improve the predictive accuracy of the physical models. However, it should be noted that when injection occurs near buried faults, none of the existing models are reliable, which is a limitation that must be addressed through both model improvement and risk avoidance.

7. Induced-Seismicity Management

The traffic-light system (TLS) is presently the primary risk-management protocol for Enhanced Geothermal System (EGS) projects. Originally formalized by Bommer et al. [70], the TLS establishes site-specific thresholds (e.g., magnitude, peak ground velocity) based on regulatory requirements or local seismic risk tolerance. When a threshold is exceeded, predefined operational responses—such as reducing or ceasing injection—are triggered automatically. Although the TLS has proven effective in several campaigns, its intrinsically reactive nature has allowed threshold breaches to occur, notably at Basel (Switzerland) and Pohang (Republic of Korea), where an M_w 3.2 event was recorded during injection.

Current improvements of the TLS focus on two complementary directions: (1) Prognostic risk assessment. Adaptive traffic-light systems (ATLS) [71] integrate forward-looking seismic forecasts to adjust thresholds and operational parameters in real time. These systems require dense, high-quality datasets. The Finnish ST1 project evaluated such forward models’ post-stimulation and concluded that existing algorithms tend to over-predict hazard; recalibration against site-specific data remains essential before operational deployment. (2) Operational mitigation measures. A suite of engineering strategies has been proposed and field-tested: gradual rate ramp-up and stepped shut-down [59], cyclic soft stimulation [72,73], fracture “caging” [74], and controlled fault creep [75].

Cyclic soft stimulation has demonstrated consistent reduction in maximum magnitude across laboratory, mine-scale, and full-field trials (e.g., the fourth stimulation stage at Pohang). Comparable results were obtained from non-seismicity-targeted cyclic treatments at Groß Schönebeck (Germany) [76] and ST1 (Finland), suggesting that alternative injection schedules may become pivotal in regions with stringent seismic constraints. Building on the generalized hydraulic-fracturing TLS framework proposed by Walters et al. [7], a real-time feedback protocol that integrates dynamic risk metrics—b-value trends, seismic-energy release efficiency (I_E), forecasted maximum magnitude, and seismogenic-plane alignment—should be implemented. This protocol enforces adjustable limits on wellhead pressure, per-stage injected volume, injection rate, and inter-stage shut-in duration (Figure 5), thereby representing a requisite evolution toward proactive, data-driven seismic-risk management.

Figure 5. Real-time management framework for induced seismicity in hot dry rock (HDR) development. Four-tier control thresholds are established through site-specific seismic risk characterization. By combining induced seismicity catalogues with injection plans, real-time prospective forecasting of induced seismic risk is conducted. Feedback-based real-time control is then implemented through measures such as limiting wellhead pressure, restricting single-injection volume, controlling injection rates, and extending injection intervals.

8. Limitations

This study conducts a cross-site comparative analysis based on publicly available data from 13 hot dry rock (HDR) projects worldwide. However, several limitations should be noted. Differences across projects—such as in monitoring network configurations, recording durations, and the level of detail in stress state descriptions—may affect the reliability of direct comparisons of seismic characteristics. Inconsistent monitoring capabilities may lead to an underestimation of the proportion of the felt earthquakes. Therefore, the 62% reported in this paper should be seen as a minimum estimate. Non-uniform magnitude scales and the use of empirical conversion formulas may introduce some uncertainty and bias into the cross-site correlation analysis between magnitude and injection parameters. Publicly available information on in situ stress magnitudes and operational strategies is limited. Although we observed a grouped linear relationship between injection pressure and maximum magnitude, caution is needed when generalizing the conclusions. Furthermore, the sample size is small, which limits its statistical power. Thus, these findings require further validation with more data.

9. Conclusions

Undoubtedly, hot dry rock (HDR) stimulations frequently generate felt events, yet destructive earthquakes remain rare. The affected volume is generally confined within 2 km of the injection well, and most projects exhibit a clear “tailing effect” and a pronounced “Kaiser effect.” Discrimination between induced and natural earthquakes relies on spatiotemporal correlations, deviations from the regional baseline, and causal links to injection parameters.

Furthermore, the dominant mechanism for felt events appears to be slip on pre-existing fractures or faults driven by pore-pressure increase and poro-elastic stress changes, with pore pressure likely serving as the primary controlling factor.

On the other hand, cross-site comparisons reveal no statistically significant correlation between total injected volume and the maximum induced magnitude. Correlations between the injection rate and magnitude seem weak; nevertheless, gradual rate modulation may reduce the likelihood of large events. The wellhead pressure appears to be a key control variable, with correlation analyses suggesting a robust positive relationship with the maximum magnitude. Low-injectivity reservoirs subjected to high wellhead pressures tend to be particularly prone to large events, whereas cyclic fracturing and similar modified injection strategies may represent important means for mitigating risk in analogous settings.

In addition, sustained low b-values, elevated seismogenic indices, and rapidly increasing seismic-energy release efficiency may serve as qualitative indicators of hazard. Within the same site, volume-based maximum-magnitude models can provide quantitative reference for evaluation, but all of models may fail when the injection occurs near major faults. In cross-site comparisons, the injection pressure seems to offer more reference value than injected volume, although further data support is needed and caution is required.

Finally, integrating comprehensive seismic risk indicators (e.g., b-value, seismic release efficiency, maximum magnitude forecasting) with key injection control parameters (e.g., wellhead pressure, single-injection volume) into a real-time control system may represent a promising research direction. Preliminary analyses suggest that a multi-metric decision framework appears worthy of further investigation for balancing induced seismic risk against reservoir stimulation performance. However, all the above evaluation metrics and the thresholds of control parameters require further validation through more field applications.