Search in the TU Delft Repository Collections
TU Delft Repository
210,241
Records
238,745
Files
81,379
Contributors
Recently Added Records
Writing software in low-level programming languages that provide full control over the memory is difficult and error-prone. Languages like C and C++ are notorious for memory-safety issues that become serious vulnerabilities. Rust guarantees memory safety at compile time through ownership, borrowing, and lifetimes, but provides an unsafe mechanism through raw pointers that alias in ways the compiler does not track. Tree Borrows is an aliasing model defining which aliasing patterns cause undefined behaviour. Miri, the official undefined behaviour detection tool for Rust, checks Tree Borrows dynamically and finds violations only on the paths it executes, but we want to detect them statically, without introducing false positives.
This thesis introduces a static and compositional analysis for detecting Tree Borrows violations in Rust, formalised on top of incorrectness separation logic (ISL) and implemented in Infer, Meta's open-source static analysis framework. We extend Infer's ISL-based checker, Pulse, to carry a Tree Borrows state in a form that allows for composing procedure summaries under bi-abduction. We achieve this by storing the trees' tags, permissions and relations as spatial resources which can combine under the separating conjunction.
We evaluate the analysis on 122 programs from Miri's borrows test suite, and it detects undefined behaviour for 19 programs that have it and reports a single false positive. Among the programs the frontend can translate, most misses, together with the introduced false positive, are caused by features of Tree Borrows that are not supported yet, such as exposed provenance and interior mutability. The analysis has been merged into the Infer repository. ...
This thesis introduces a static and compositional analysis for detecting Tree Borrows violations in Rust, formalised on top of incorrectness separation logic (ISL) and implemented in Infer, Meta's open-source static analysis framework. We extend Infer's ISL-based checker, Pulse, to carry a Tree Borrows state in a form that allows for composing procedure summaries under bi-abduction. We achieve this by storing the trees' tags, permissions and relations as spatial resources which can combine under the separating conjunction.
We evaluate the analysis on 122 programs from Miri's borrows test suite, and it detects undefined behaviour for 19 programs that have it and reports a single false positive. Among the programs the frontend can translate, most misses, together with the introduced false positive, are caused by features of Tree Borrows that are not supported yet, such as exposed provenance and interior mutability. The analysis has been merged into the Infer repository. ...
Writing software in low-level programming languages that provide full control over the memory is difficult and error-prone. Languages like C and C++ are notorious for memory-safety issues that become serious vulnerabilities. Rust guarantees memory safety at compile time through ownership, borrowing, and lifetimes, but provides an unsafe mechanism through raw pointers that alias in ways the compiler does not track. Tree Borrows is an aliasing model defining which aliasing patterns cause undefined behaviour. Miri, the official undefined behaviour detection tool for Rust, checks Tree Borrows dynamically and finds violations only on the paths it executes, but we want to detect them statically, without introducing false positives.
This thesis introduces a static and compositional analysis for detecting Tree Borrows violations in Rust, formalised on top of incorrectness separation logic (ISL) and implemented in Infer, Meta's open-source static analysis framework. We extend Infer's ISL-based checker, Pulse, to carry a Tree Borrows state in a form that allows for composing procedure summaries under bi-abduction. We achieve this by storing the trees' tags, permissions and relations as spatial resources which can combine under the separating conjunction.
We evaluate the analysis on 122 programs from Miri's borrows test suite, and it detects undefined behaviour for 19 programs that have it and reports a single false positive. Among the programs the frontend can translate, most misses, together with the introduced false positive, are caused by features of Tree Borrows that are not supported yet, such as exposed provenance and interior mutability. The analysis has been merged into the Infer repository.
This thesis introduces a static and compositional analysis for detecting Tree Borrows violations in Rust, formalised on top of incorrectness separation logic (ISL) and implemented in Infer, Meta's open-source static analysis framework. We extend Infer's ISL-based checker, Pulse, to carry a Tree Borrows state in a form that allows for composing procedure summaries under bi-abduction. We achieve this by storing the trees' tags, permissions and relations as spatial resources which can combine under the separating conjunction.
We evaluate the analysis on 122 programs from Miri's borrows test suite, and it detects undefined behaviour for 19 programs that have it and reports a single false positive. Among the programs the frontend can translate, most misses, together with the introduced false positive, are caused by features of Tree Borrows that are not supported yet, such as exposed provenance and interior mutability. The analysis has been merged into the Infer repository.
Student report
(2026)
-
D. Beeren, T.C. van Esch, D.A.G. van Krimpen, M. Sijsling, B. Vink, Roger Pacheco-Castro, José A. Á. Antolínez, Paul A. Korswagen, M.A. Cabrera, Jaime Arriaga, Alec Torres Freyermuth
San Felipe is a low lying coastal town on the northern coast of the Yucatán Peninsula, bordered by a lagoon and surrounded by wetlands. The town is repeatedly flooded by hurricanes, tropical storms and cold fronts (Nortes). This report investigates the cause and impact of flooding in San Felipe and identifies the critical drivers behind it. A system analysis of the coastal, geohydrological and urban system is combined with an analysis of historical data and fieldwork, consisting of interviews with the local community and GPS measurements. Waves are modelled with SWAN, groundwater flow with MODFLOW and overland flooding with SFINCS. The contribution of precipitation, storm surge and initial infiltration is quantified with a 2^3 factorial design of eight SFINCS simulations, using 50-year return period values for precipitation (306 mm in three days) and storm surge (0.87 m). Finally, the SFINCS simulation of the 2024 Norte event is used to assess the impact of the flood event on the building structures of San Felipe.
Storm surge is the primary driver for the total flooded area (51.8%) and for the average maximum inundation depth (87.4%), whereas precipitation is the primary control on the inundation duration (81.5%). The drivers amplify each other non-linearly, while the infiltration capacity dampens the maximum inundation depths and the duration of the floods. Waves have a significant influence on flooding in San Felipe, but are not yet included in the compound flooding analysis.The MODFLOW model shows that the areas along the quay walls in the north, the main street and the south east of the town are most prone to flooding. This corresponds largely with the experiences of the local residents, but not all critical locations are captured, which indicates that groundwater is not the only driver. The structural vulnerability analysis indicates that timber buildings with a low floor level and no additional protection are the most susceptible to flood damage, whereas more elevated concrete and masonry structures with a protective finish are the least susceptible. Combined with the flood hazard characteristics of the simulated February 2024 event the results suggest that buildings located in the areas bordering the mangroves, the quay wall and the low-lying main street are expected to experience the greatest damage during flood events.
...
Storm surge is the primary driver for the total flooded area (51.8%) and for the average maximum inundation depth (87.4%), whereas precipitation is the primary control on the inundation duration (81.5%). The drivers amplify each other non-linearly, while the infiltration capacity dampens the maximum inundation depths and the duration of the floods. Waves have a significant influence on flooding in San Felipe, but are not yet included in the compound flooding analysis.The MODFLOW model shows that the areas along the quay walls in the north, the main street and the south east of the town are most prone to flooding. This corresponds largely with the experiences of the local residents, but not all critical locations are captured, which indicates that groundwater is not the only driver. The structural vulnerability analysis indicates that timber buildings with a low floor level and no additional protection are the most susceptible to flood damage, whereas more elevated concrete and masonry structures with a protective finish are the least susceptible. Combined with the flood hazard characteristics of the simulated February 2024 event the results suggest that buildings located in the areas bordering the mangroves, the quay wall and the low-lying main street are expected to experience the greatest damage during flood events.
...
San Felipe is a low lying coastal town on the northern coast of the Yucatán Peninsula, bordered by a lagoon and surrounded by wetlands. The town is repeatedly flooded by hurricanes, tropical storms and cold fronts (Nortes). This report investigates the cause and impact of flooding in San Felipe and identifies the critical drivers behind it. A system analysis of the coastal, geohydrological and urban system is combined with an analysis of historical data and fieldwork, consisting of interviews with the local community and GPS measurements. Waves are modelled with SWAN, groundwater flow with MODFLOW and overland flooding with SFINCS. The contribution of precipitation, storm surge and initial infiltration is quantified with a 2^3 factorial design of eight SFINCS simulations, using 50-year return period values for precipitation (306 mm in three days) and storm surge (0.87 m). Finally, the SFINCS simulation of the 2024 Norte event is used to assess the impact of the flood event on the building structures of San Felipe.
Storm surge is the primary driver for the total flooded area (51.8%) and for the average maximum inundation depth (87.4%), whereas precipitation is the primary control on the inundation duration (81.5%). The drivers amplify each other non-linearly, while the infiltration capacity dampens the maximum inundation depths and the duration of the floods. Waves have a significant influence on flooding in San Felipe, but are not yet included in the compound flooding analysis.The MODFLOW model shows that the areas along the quay walls in the north, the main street and the south east of the town are most prone to flooding. This corresponds largely with the experiences of the local residents, but not all critical locations are captured, which indicates that groundwater is not the only driver. The structural vulnerability analysis indicates that timber buildings with a low floor level and no additional protection are the most susceptible to flood damage, whereas more elevated concrete and masonry structures with a protective finish are the least susceptible. Combined with the flood hazard characteristics of the simulated February 2024 event the results suggest that buildings located in the areas bordering the mangroves, the quay wall and the low-lying main street are expected to experience the greatest damage during flood events.
Storm surge is the primary driver for the total flooded area (51.8%) and for the average maximum inundation depth (87.4%), whereas precipitation is the primary control on the inundation duration (81.5%). The drivers amplify each other non-linearly, while the infiltration capacity dampens the maximum inundation depths and the duration of the floods. Waves have a significant influence on flooding in San Felipe, but are not yet included in the compound flooding analysis.The MODFLOW model shows that the areas along the quay walls in the north, the main street and the south east of the town are most prone to flooding. This corresponds largely with the experiences of the local residents, but not all critical locations are captured, which indicates that groundwater is not the only driver. The structural vulnerability analysis indicates that timber buildings with a low floor level and no additional protection are the most susceptible to flood damage, whereas more elevated concrete and masonry structures with a protective finish are the least susceptible. Combined with the flood hazard characteristics of the simulated February 2024 event the results suggest that buildings located in the areas bordering the mangroves, the quay wall and the low-lying main street are expected to experience the greatest damage during flood events.
Combining spatially separated frequency-modulated continuous-wave (FMCW) radar nodes can improve angular resolution beyond that of a single module. Coherent use of this distributed aperture requires carrier-phase calibration as well as timing and frequency synchronization. This thesis evaluates the localization sensitivity of two independently clocked 77 GHz radar nodes forming two monostatic and two reciprocal bistatic apertures. A common processing framework compares Bartlett beamforming, independent FOCUSS, Block-FOCUSS, multi-task Bayesian compressive sensing (MT-BCS), and coherent FOCUSS, alongside coherent and non-coherent Cartesian backpropagation. The contribution is an estimator-specific assessment linking residual phase structure after synchronization to angular accuracy, source-count recovery, and relative-separation success. Controlled simulations distinguish intra-aperture manifold distortion, inter-aperture block rotations, and phase evolution within a coherent processing interval (CPI).
In the static-phase studies, MT-BCS gives the most consistent sparse close-pair performance without requiring a common carrier phase. Coherent processing produces finer ideal spatial responses but is more sensitive to calibration and baseline-dependent ambiguity. In single-target CPI tests with fixed waveform settings, peak-tracked coherent FOCUSS with a frozen reference retains angular root-mean-square error (RMSE) below 0.05° for all tested residual frequency offsets through 100 Hz. Strong random phase walks and incompatible target/reference Doppler readouts can degrade coherent estimates; accurate dominant-peak angles can coexist with incorrect source counts.
The main physical campaign uses a 25 cm radar baseline, nine target separations, and 72 frame-level cases. Some estimators recover relative spacing, but none achieves absolute-angle success: exactly two targets with both bearing errors within 2°. Later 14.03 and 19.5 cm baseline campaigns compare processing without compensation against a separate spherical reference. Reference compensation remains limited by geometric uncertainty, while scene-assisted correction is an in-sample diagnostic. A measurement-matched simulator reproduces selected range–Doppler structures but not the measured estimator rankings. Together, these results identify when residual phase permits useful distributed-aperture localization and why a narrow coherent response alone is insufficient evidence of accurate physical localization. ...
In the static-phase studies, MT-BCS gives the most consistent sparse close-pair performance without requiring a common carrier phase. Coherent processing produces finer ideal spatial responses but is more sensitive to calibration and baseline-dependent ambiguity. In single-target CPI tests with fixed waveform settings, peak-tracked coherent FOCUSS with a frozen reference retains angular root-mean-square error (RMSE) below 0.05° for all tested residual frequency offsets through 100 Hz. Strong random phase walks and incompatible target/reference Doppler readouts can degrade coherent estimates; accurate dominant-peak angles can coexist with incorrect source counts.
The main physical campaign uses a 25 cm radar baseline, nine target separations, and 72 frame-level cases. Some estimators recover relative spacing, but none achieves absolute-angle success: exactly two targets with both bearing errors within 2°. Later 14.03 and 19.5 cm baseline campaigns compare processing without compensation against a separate spherical reference. Reference compensation remains limited by geometric uncertainty, while scene-assisted correction is an in-sample diagnostic. A measurement-matched simulator reproduces selected range–Doppler structures but not the measured estimator rankings. Together, these results identify when residual phase permits useful distributed-aperture localization and why a narrow coherent response alone is insufficient evidence of accurate physical localization. ...
Combining spatially separated frequency-modulated continuous-wave (FMCW) radar nodes can improve angular resolution beyond that of a single module. Coherent use of this distributed aperture requires carrier-phase calibration as well as timing and frequency synchronization. This thesis evaluates the localization sensitivity of two independently clocked 77 GHz radar nodes forming two monostatic and two reciprocal bistatic apertures. A common processing framework compares Bartlett beamforming, independent FOCUSS, Block-FOCUSS, multi-task Bayesian compressive sensing (MT-BCS), and coherent FOCUSS, alongside coherent and non-coherent Cartesian backpropagation. The contribution is an estimator-specific assessment linking residual phase structure after synchronization to angular accuracy, source-count recovery, and relative-separation success. Controlled simulations distinguish intra-aperture manifold distortion, inter-aperture block rotations, and phase evolution within a coherent processing interval (CPI).
In the static-phase studies, MT-BCS gives the most consistent sparse close-pair performance without requiring a common carrier phase. Coherent processing produces finer ideal spatial responses but is more sensitive to calibration and baseline-dependent ambiguity. In single-target CPI tests with fixed waveform settings, peak-tracked coherent FOCUSS with a frozen reference retains angular root-mean-square error (RMSE) below 0.05° for all tested residual frequency offsets through 100 Hz. Strong random phase walks and incompatible target/reference Doppler readouts can degrade coherent estimates; accurate dominant-peak angles can coexist with incorrect source counts.
The main physical campaign uses a 25 cm radar baseline, nine target separations, and 72 frame-level cases. Some estimators recover relative spacing, but none achieves absolute-angle success: exactly two targets with both bearing errors within 2°. Later 14.03 and 19.5 cm baseline campaigns compare processing without compensation against a separate spherical reference. Reference compensation remains limited by geometric uncertainty, while scene-assisted correction is an in-sample diagnostic. A measurement-matched simulator reproduces selected range–Doppler structures but not the measured estimator rankings. Together, these results identify when residual phase permits useful distributed-aperture localization and why a narrow coherent response alone is insufficient evidence of accurate physical localization.
In the static-phase studies, MT-BCS gives the most consistent sparse close-pair performance without requiring a common carrier phase. Coherent processing produces finer ideal spatial responses but is more sensitive to calibration and baseline-dependent ambiguity. In single-target CPI tests with fixed waveform settings, peak-tracked coherent FOCUSS with a frozen reference retains angular root-mean-square error (RMSE) below 0.05° for all tested residual frequency offsets through 100 Hz. Strong random phase walks and incompatible target/reference Doppler readouts can degrade coherent estimates; accurate dominant-peak angles can coexist with incorrect source counts.
The main physical campaign uses a 25 cm radar baseline, nine target separations, and 72 frame-level cases. Some estimators recover relative spacing, but none achieves absolute-angle success: exactly two targets with both bearing errors within 2°. Later 14.03 and 19.5 cm baseline campaigns compare processing without compensation against a separate spherical reference. Reference compensation remains limited by geometric uncertainty, while scene-assisted correction is an in-sample diagnostic. A measurement-matched simulator reproduces selected range–Doppler structures but not the measured estimator rankings. Together, these results identify when residual phase permits useful distributed-aperture localization and why a narrow coherent response alone is insufficient evidence of accurate physical localization.
Dense retrievers are often trained using contrastive learning to differentiate a positive document from a negative one. The harder the negative, the better the training signal and thus the better the distinction. In-batch training takes its negatives from other queries in the same batch, while cross-batch increases the negative document size by using several batches of them. These documents, however, are most of the time easy to tell apart, and larger batches come at the cost of more GPU memory. ANCE finds harder negatives by periodically re-encoding the full corpus with a new checkpoint of the model, which is expensive to do on limited hardware. This thesis investigates whether hard negatives can be found without re-encoding the whole corpus. GRASS is then proposed, which mines documents from a small cache that periodically gets updated and refreshed, and uses the model’s uncertainty in the selection of hard negatives. After fine-tuning the pretrained retriever BGE-M3, GRASS scored on par with ANCE on the 12 BRIGHT domains and higher than cross-batch and in-batch, while each mining round took about a third of the time of an ANCE round.
...
Dense retrievers are often trained using contrastive learning to differentiate a positive document from a negative one. The harder the negative, the better the training signal and thus the better the distinction. In-batch training takes its negatives from other queries in the same batch, while cross-batch increases the negative document size by using several batches of them. These documents, however, are most of the time easy to tell apart, and larger batches come at the cost of more GPU memory. ANCE finds harder negatives by periodically re-encoding the full corpus with a new checkpoint of the model, which is expensive to do on limited hardware. This thesis investigates whether hard negatives can be found without re-encoding the whole corpus. GRASS is then proposed, which mines documents from a small cache that periodically gets updated and refreshed, and uses the model’s uncertainty in the selection of hard negatives. After fine-tuning the pretrained retriever BGE-M3, GRASS scored on par with ANCE on the 12 BRIGHT domains and higher than cross-batch and in-batch, while each mining round took about a third of the time of an ANCE round.
Monitoring climate change adaptation (CCA) in the heritage sector, especially in the context of World Heritage (WH) management, requires systematic analyses of policy documents for CCA strategies, a process that is difficult to scale. We ask whether automated retrieval systems can prioritise expert annotation effort: given a CCA concept, can a system rank passages of a management plan so that those an expert would annotate appear near the top, without any in-domain supervision? This objective is formulated as an information retrieval problem with a new benchmark dataset on climate-adaptive heritage management strategies, created from expert-labelled text chunks in official Management Plans of 11 Dutch WH properties. Four retrieval approaches (BM25, a bi-encoder, a dense retriever with reranker, and a cross-encoder reranker) are evaluated and compared. The cross-encoder reranker achieves the strongest performance across labels. An LLM-as-judge evaluation further shows that top-ranked unlabelled chunks could be relevant, suggesting that the existing annotations are precise but not exhaustive. Per-label analysis identifies which concepts are automatically retrievable and which remain beyond the reach of general-purpose retrieval. Our benchmark establishes a foundation for scalable mapping and monitoring of climate change adaptation in the context of heritage management.
...
Monitoring climate change adaptation (CCA) in the heritage sector, especially in the context of World Heritage (WH) management, requires systematic analyses of policy documents for CCA strategies, a process that is difficult to scale. We ask whether automated retrieval systems can prioritise expert annotation effort: given a CCA concept, can a system rank passages of a management plan so that those an expert would annotate appear near the top, without any in-domain supervision? This objective is formulated as an information retrieval problem with a new benchmark dataset on climate-adaptive heritage management strategies, created from expert-labelled text chunks in official Management Plans of 11 Dutch WH properties. Four retrieval approaches (BM25, a bi-encoder, a dense retriever with reranker, and a cross-encoder reranker) are evaluated and compared. The cross-encoder reranker achieves the strongest performance across labels. An LLM-as-judge evaluation further shows that top-ranked unlabelled chunks could be relevant, suggesting that the existing annotations are precise but not exhaustive. Per-label analysis identifies which concepts are automatically retrievable and which remain beyond the reach of general-purpose retrieval. Our benchmark establishes a foundation for scalable mapping and monitoring of climate change adaptation in the context of heritage management.