kth.sePublications KTH
Change search
CiteExportLink to record
Permanent link

Direct link
Cite
Citation style
  • apa
  • ieee
  • modern-language-association-8th-edition
  • vancouver
  • Other style
More styles
Language
  • de-DE
  • en-GB
  • en-US
  • fi-FI
  • nn-NO
  • nn-NB
  • sv-SE
  • Other locale
More languages
Output format
  • html
  • text
  • asciidoc
  • rtf
A Computational Study on the Catalytic Adsorption of Gaseous Probes on Titanium Dioxide Nanoclusters
KTH, School of Engineering Sciences in Chemistry, Biotechnology and Health (CBH), Chemistry, Applied Physical Chemistry.
2026 (English)Independent thesis Advanced level (degree of Master (Two Years)), 20 credits / 30 HE creditsStudent thesis
Abstract [sv]

Atomistiska simuleringar är avgörande för att avkoda kemiska fenomen i gränsskikt, men traditionell modellering innebär en viktig avvägning mellan den höga kvantmekaniska noggrannheten hos densitetsfunktionalteori (DFT) och beräkningseffektiviteten hos empiriska kraftfält. För att överbrygga detta gap implementerar denna avhandling ett systematiskt ramverk för benchmarking och verifiering som använder universella maskininlärda interatomiska potentialer (MLIP), SevenNet-Omni, tillsammans med lokala DFT-beräkningsspår. Med fokus på titandioxid TiO2 nanokluster identifieras lokaliserade ytaktiva platser via den maximala elektrostatiska potentialen VS,max med hjälp av Molecular Surface Properties Approach (MSPA) på PBE0/def2-TZVP-nivån. Fyra representativa Lewis-bas-probmolekyler – vatten (H2O), vätesulfid (H2S), ammoniak (NH3) och kolmonoxid (CO) – utsätts därefter för systematisk analys med hjälp av snabba MLIP-kalkylatorer för att screena för termodynamiska minimum-konfigurationer. De strukturella och energetiska resultaten avslöjar MLIPs svagheter och styrkor. Linjära regressionsutvärderingar mot referenshybrid-DFT-funktionaler visar hög korrelation (R2 > 0.95), där den hybridtränade omol25_low-potentialen ger överlägsen strukturell relaxationsfidelitet, framgångsrikt fångar snäva, kontraherade klusterramverk och reproducerar relativa energier med trippel-zeta-nivåprecision. Den semilokala matpes_r2scan-potentialen underpresterar geometriskt men uppvisar utmärkt noggrannhet i att förutsäga nativa termodynamiska adsorptionsenergi-variationer. För att utnyttja dessa komplementära egenskaper föreslås en MLIPopt + DFT_sp hybridpipeline, där strukturer som är föroptimerade via SevenNet-Omni skickas direkt till högnivå-energiberäkningar (triple-zeta). Detta hybridarbetsflöde ger adsorptionsenergier som korrelerar utmärkt med standard DFT-beräkningar samtidigt som förkortar beräkningstiden med flera storleksordningar (från nästan åtta timmar ner till ungefär en minut). I slutändan bekräftar detta arbete att universella maskininlärningspotentialer kan fungera som exakta strukturella deskriptorer för metalloxidgränssnitt, och erbjuder ett mycket skalbart och universellt alternativ för att accelerera framtida högkapacitetsscreening inom heterogen katalys. 

Abstract [en]

Atomistic simulations are essential for decoding interfacial chemical phenomena, yet traditional modeling faces a severe trade-off between the high quantum-mechanical accuracy of Density Functional Theory (DFT) and the computational efficiency of empirical force fields. To bridge this gap, this thesis implements a systematic performance benchmarking and verification framework utilizing the cross-domain universal Machine Learning Interatomic Potential (MLIP) architecture, SevenNet-Omni, alongside conventional DFT computational tracks. Focused on titanium dioxide (TiO2) nanoclusters, localized surface active sites are identified via the maximum electrostatic potential (VS,max) using the Molecular Surface Properties Approach (MSPA) at the PBE0/def2-TZVP level. Four representative Lewis base probe molecules—water (H2O), hydrogen sulfide (H2S), ammonia (NH3), and carbon monoxide (CO})—are subsequently subjected to systematic orientation sampling using fast MLIP calculators to screen for thermodynamic minimum configurations.The structural and energetic benchmarks reveal distinct performance profiles across the multi-task MLIP channels. Linear regression evaluations against reference hybrid DFT functionals demonstrate exceptional correlation (R2 > 0.95), where the hybrid-trained omol25_low potential delivers superior structural relaxation fidelity, successfully capturing tight, contracted cluster frameworks and reproducing relative energy splits with triple-zeta level precision. Conversely, the semilocal matpes_r2scan potential underperforms geometrically but exhibits excellent accuracy in predicting native thermodynamic adsorption energy variations. To leverage these complementary traits, a tiered MLIPopt + DFT_sp hybride pipeline is proposed, where structures pre-optimized via SevenNet-Omni are passed directly to high-level triple-zeta single-point evaluations. This hybrid workflow yields single-point adsorption energies that correlate excellently with standard all-DFT relaxation tracks while shortening the computational time by several orders of magnitude (from nearly eight hours down to approximately one minute). Ultimately, this work validates that universal machine learning potentials can serve as accurate structural descriptors for metal oxide interfaces, offering a highly scalable and universal alternative to accelerate future high-throughput screening in heterogeneous catalysis.

Place, publisher, year, edition, pages
2026.
Series
TRITA-CBH-GRU ; 2026:279
Keywords [en]
machine learning interatomic potential, molecular surface properties approach, density functional theory, catalytic adsorption, titanium dioxide clusters
Keywords [sv]
maskininlärda interatomiska potentialer, molekylära ytegenskaper, täthetsfunktionalteori, katalytisk adsorption, titandioxid-kluster
National Category
Theoretical Chemistry
Identifiers
URN: urn:nbn:se:kth:diva-384169OAI: oai:DiVA.org:kth-384169DiVA, id: diva2:2080052
Subject / course
Chemistry
Educational program
Degree of Master - Molecular Science and Engineering
Examiners
Available from: 2026-06-26 Created: 2026-06-26

Open Access in DiVA

fulltext(2643 kB)94 downloads
File information
File name FULLTEXT01.pdfFile size 2643 kBChecksum SHA-512
82dcda419a961a4ce9c31a292b8a6be20b99b4595f0e21e648abdedf46f3520ca947770bf306a2e6f22dbb6994e35bf08afb521950e351bb77327b412308c592
Type fulltextMimetype application/pdf

By organisation
Applied Physical Chemistry
Theoretical Chemistry

Search outside of DiVA

GoogleGoogle Scholar
Total: 94 downloads
The number of downloads is the sum of all downloads of full texts. It may include eg previous versions that are now no longer available

urn-nbn

Altmetric score

urn-nbn
Total: 949 hits
CiteExportLink to record
Permanent link

Direct link
Cite
Citation style
  • apa
  • ieee
  • modern-language-association-8th-edition
  • vancouver
  • Other style
More styles
Language
  • de-DE
  • en-GB
  • en-US
  • fi-FI
  • nn-NO
  • nn-NB
  • sv-SE
  • Other locale
More languages
Output format
  • html
  • text
  • asciidoc
  • rtf