125878 research outputs found
Sort by
STEREOSELECTIVE (CO)POLYMERIZATION OF BIO-BASED CONJUGATED DIENES DERIVED FROM NATURAL ALDEHYDES USING [OSSO]-TITANIUM CATALYSTS
No full textTHE TRANSITION TOWARD A SUSTAINABLE POLYMER INDUSTRY REQUIRES THE DESIGN OF NEW MATERIALS DERIVED FROM RENEWABLE RESOURCES, CAPABLE OF COMBINING HIGH PERFORMANCE WITH ENVIRONMENTAL RESPONSIBILITY. IN THIS CONTEXT, THIS WORK FOCUSES ON THE SYNTHESIS AND STEREOSELECTIVE POLYMERIZATION OF NOVEL BIO-BASED CONJUGATED DIENES OBTAINED FROM NATURAL ALDEHYDES SUCH AS CINNAMALDEHYDE, PERILLALDEHYDE, AND CITRAL. SPECIFICALLY, THREE MONOMERS, TRANS-1-PHENYL-1,3-BUTADIENE (1PB), S-4-ISOPROPENYL-1-VINYL-1-CYCLOHEXENE (IVC), AND (E)-4,8-DIMETHYL-1,3,7-NONATRIENE (DMNT), WERE SYNTHESIZED AND SUBSEQUENTLY POLYMERIZED UNDER OPTIMIZED CONDITIONS USING TITANIUM [OSSO]-TYPE COMPLEXES BEARING DIFFERENT SUBSTITUENTS ON THE AROMATIC RINGS OF THE LIGAND BACKBONE.
ALL POLYMERIZATIONS AFFORDED HIGHLY REGIO- AND STEREOREGULAR POLYMERS WITH REMARKABLE ISOTACTICITY, HIGH YIELDS, AND SIGNIFICANT MOLECULAR WEIGHTS. THE RESULTING POLYMERS WERE THOROUGHLY CHARACTERIZED BY NMR SPECTROSCOPY AS WELL AS THERMAL AND MECHANICAL ANALYSES TO ELUCIDATE THE RELATIONSHIPS BETWEEN MICROSTRUCTURE AND PROPERTIES. MOREOVER, THE MONOMERS WERE COPOLYMERIZED EITHER AMONG THEMSELVES OR WITH TWO LINEAR TERPENES, Β-MYRCENE AND Β-OCIMENE, TO OBTAIN FULLY RENEWABLE COPOLYMERS EXHIBITING TUNABLE ELASTOMERIC BEHAVIOR. THESE COPOLYMERIZATIONS WERE PERFORMED EMPLOYING THE SAME CATALYTIC SYSTEMS UNDER THE OPTIMAL CONDITIONS IDENTIFIED FOR THE HOMOPOLYMERIZATIONS.
OWING TO THE PRESENCE OF RESIDUAL CARBON-CARBON DOUBLE BONDS IN THE POLYMER BACKBONE, POLY(IVC) WAS USED AS A MODEL POLYMER TO INVESTIGATE BOTH REVERSIBLE AND IRREVERSIBLE CROSS-LINKING REACTIONS AIMED AT ENHANCING MECHANICAL PERFORMANCE. IN ADDITION, KINETIC STUDIES AND MONOMER REACTIVITY ORDER ANALYSES WERE CONDUCTED FOR EACH HOMO- AND COPOLYMERIZATION SYSTEM THROUGH IN SITU NMR TECHNIQUES, PROVIDING INSIGHT INTO THE MECHANISTIC ASPECTS OF THE POLYMERIZATION PROCESSES.
OVERALL, THIS RESEARCH DEMONSTRATES THE POTENTIAL OF TITANIUM [OSSO]-CATALYZED STEREOSELECTIVE POLYMERIZATION AS A POWERFUL STRATEGY FOR CONVERTING NATURAL BUILDING BLOCKS INTO HIGH-PERFORMANCE, RENEWABLE POLYMERS, PAVING THE WAY TOWARD SUSTAINABLE ELASTOMERIC MATERIALS WITH CONTROLLED MICROSTRUCTURES AND TAILORED PROPERTIES
An Integrated Assessment Model for Evaluating Motor Skills in Trainee Primary School Teachers
No full textA MULTIDISCIPLINARY APPROACH FOR THE STUDY OF ORIGANUM GENUS
No full textTHIS STUDY INVESTIGATES THE PHYTOCHEMICAL PROFILE AND BIOLOGICAL POTENTIAL OF SELECTED SPECIES OF THE GENUS ORIGANUM, WITH A FOCUS ON O. DICTAMNUS, O. MAJORANA, O. VULGARE, AND O. HERACLEOTICUM. THE WORK WAS STRUCTURED INTO THREE MAIN OBJECTIVES. FIRST, THE VOLATILE AND NON-VOLATILE FRACTIONS OF O. DICTAMNUS WERE EVALUATED. THE ESSENTIAL OIL WAS CHARACTERISED THROUGH GC–MS AND EVALUATED FOR ITS INHIBITORY EFFECTS ON KEY ENZYMES INVOLVED IN NEURODEGENERATIVE AND METABOLIC DISORDERS. KINETIC ANALYSES WERE PERFORMED TO CLARIFY THE MODE OF INHIBITION AND BY IN SILICO ANALYSIS, MOLECULAR DOCKING SUPPORTED THE IN VITRO RESULTS BY CLARIFYING THE INTERACTIONS OF THE MAIN CONSTITUENTS, IN PARTICULAR CARVACROL AND P-CYMENE, WITH THE ACTIVE SITES OF THE ENZYME. FOR THE NON-VOLATILE FRACTION, THREE EXTRACTION METHODS (MACERATION WITH 70% ETHANOL, MACERATION WITH 20% ETHANOL, AND PRESSURIZED LIQUID EXTRACTION (PLE) WITH 20% ETHANOL) WERE COMPARED IN TERMS OF EXTRACTION EFFICIENCY AND BIOLOGICAL ACTIVITY (ANTIOXIDANT AND ENZYMATIC ASSAYS). THE EXTRACT OBTAINED VIA PLE, WHICH SHOWED THE BEST OVERALL PERFORMANCE, WAS FURTHER ANALYZED BY LC-HRESIMS/MS TO CHARACTERIZE ITS METABOLIC PROFILE. THE SECOND OBJECTIVE CONSISTED OF AN INTERSPECIFIC COMPARISON OF ESSENTIAL OILS FROM O. MAJORANA, O. VULGARE, AND O. HERACLEOTICUM. THEIR CHEMICAL COMPOSITION WAS DETERMINED THROUGH GC–MS, AND THEIR ANTIOXIDANT PROPERTIES AND ENZYME INHIBITORY ACTIVITIES WERE ASSESSED ON TARGETS RELATED TO NEURODEGENERATIVE AND METABOLIC DISEASES. ENZYME KINETIC STUDIES WERE PERFORMED TO INVESTIGATE POTENTIAL DIFFERENCES IN THEIR MECHANISMS OF ACTION. THE THIRD OBJECTIVE FOCUSED ON AN INTRASPECIFIC COMPARISON OF O. HERACLEOTICUM COLLECTED FROM DIFFERENT ITALIAN REGIONS. EIGHT SAMPLES WERE ANALYSED FOR THE VOLATILE FRACTION, ALLOWING THE IDENTIFICATION OF CHEMICAL VARIABILITY THROUGH GC–MS AND MULTIVARIATE STATISTICAL ANALYSIS. SELECTED SAMPLES WERE TESTED FOR ACTIVITIES RELATED TO THE CENTRAL NERVOUS SYSTEM AND METABOLIC DISORDERS. FOR THE NON-VOLATILE FRACTION, SIX SAMPLES WERE EXTRACTED USING THE OPTIMIZED PLE METHOD AND EVALUATED FOR THEIR CHEMICAL FEATURES AND BIOLOGICAL PROPERTIES, PARTICULARLY REGARDING ENZYMES INVOLVED IN METABOLIC DYSFUNCTION. OVERALL, THE FINDINGS DEMONSTRATE CHEMICAL DIVERSITY AMONG ORIGANUM SPECIES AND HIGHLIGHT THEIR POTENTIAL AS SOURCES OF NATURAL INHIBITORS TARGETING ENZYMES IMPLICATED IN NEURODEGENERATIVE AND METABOLIC DISORDERS. THE STUDY SUPPORTS THE RELEVANCE OF ORIGANUM SPP. AS PROMISING CANDIDATES FOR NUTRACEUTICAL AND PHYTOTHERAPEUTIC APPLICATION
Fast Quantum-Mechanical Predictions of Radiative and Non-Radiative Rates in Organic Optoelectronics
No full textPROGETTAZIONE DI ACCELERATORI HARDWARE PER RETI NEURALI MINIATURIZZATI E A ULTRA-BASSO CONSUMO PER L’IN-SENSOR COMPUTING
No full textLA RAPIDA DIFFUSIONE DEI DISPOSITIVI IOT PONE LA NECESSITÀ DI INTEGRARE CAPACITÀ DI INTELLIGENZA DIRETTAMENTE A LIVELLO DI SENSORE, GARANTENDO AL CONTEMPO ELEVATA ACCURATEZZA E CONSUMI ENERGETICI ESTREMAMENTE RIDOTTI. QUESTA TESI DI DOTTORATO ANALIZZA IL PARADIGMA DELL’IN-SENSOR COMPUTING (ISC) ATTRAVERSO UNA METODOLOGIA DI CO-PROGETTAZIONE HARDWARE–SOFTWARE GUIDATA DAI VINCOLI HARDWARE, FINALIZZATA ALLA REALIZZAZIONE DI ACCELERATORI NEURALI MINIATURIZZATI E A BASSISSIMA POTENZA, STRETTAMENTE INTEGRATI CON SENSORI MEMS. L’APPROCCIO PROPOSTO DEFINISCE FIN DALLE PRIME FASI DI PROGETTO LA TOPOLOGIA DELLE RETI NEURALI, LE STRATEGIE DI QUANTIZZAZIONE E L’ARITMETICA A PUNTO FISSO, AL FINE DI RISPETTARE STRINGENTI VINCOLI DI AREA, POTENZA E MEMORIA. LE SOLUZIONI SVILUPPATE SONO VALIDATE MEDIANTE PROTOTIPAZIONE SU FPGA E SINTESI CMOS.
LA VALIDITÀ DELLA METODOLOGIA È DIMOSTRATA ATTRAVERSO TRE CASI DI STUDIO APPLICATIVI, SVILUPPATI IN COLLABORAZIONE CON STMICROELECTRONICS. IL PRIMO RIGUARDA L’ELABORAZIONE AUDIO PER APPLICAZIONI DI KEYWORD SPOTTING, IN CUI UNA RETE NEURALE CONVOLUZIONALE 1D SOSTITUISCE LA TRADIZIONALE CATENA CIC+FIR PER LA CONVERSIONE PDM–PCM, INTEGRANDO FILTRAGGIO E DECIMAZIONE IN UN UNICO BLOCCO NEURALE. LA SOLUZIONE PRODUCE SEGNALI PCM A 8 BIT E 16 KHZ CON UN RAPPORTO SEGNALE-RUMORE DI 48 DB, MANTENENDO UN’ACCURATEZZA COMPLESSIVA DELL’89%. IL CORE SINTETIZZATO IN TECNOLOGIA CMOS A 130 NM RAGGIUNGE UNA POTENZA DI 128,7 ΜW/MHZ IN UN’AREA INFERIORE A 1 MM^2.
IL SECONDO CASO DI STUDIO AFFRONTA LA MANUTENZIONE PREDITTIVA BASATA SU SEGNALI VIBRAZIONALI MEDIANTE UNA PIPELINE IBRIDA ED EVENT-DRIVEN, CHE COMBINA UN AUTOENCODER IN-SENSOR SEMPRE ATTIVO, PARZIALMENTE BINARIZZATO, PER IL RILEVAMENTO DI ANOMALIE, CON UN CLASSIFICATORE SU MICROCONTROLLORE ATTIVATO SU RICHIESTA. L’APPROCCIO CONSENTE DI OTTENERE PRESTAZIONI DI RILEVAMENTO DI ANOMALIE PROSSIME ALLO STATO DELL’ARTE (AUC PARI A 0,99) E UN’ACCURATEZZA DI CLASSIFICAZIONE FINO AL 94,83%, SOSTENENDO RATE DI DATI IN USCITA DAL SENSORE FINO A 365 KHZ. I RISULTATI DI SINTESI RIPORTANO UN’AREA DI 0,49 MM^2 IN TECNOLOGIA CMOS A 65 NM E UNA POTENZA DINAMICA DI 138,6 ΜW/MHZ.
IL TERZO CASO DI STUDIO È DEDICATO ALLA COMPENSAZIONE DELLO STRESS TERMICO NEI SENSORI DI PRESSIONE MEMS. VIENE PROPOSTA UN’UNITÀ DI COMPENSAZIONE RICONFIGURABILE BASATA SU INTELLIGENZA ARTIFICIALE (AI-RESCU), CHE COMBINA UN MECCANISMO DI TRIGGER ADATTIVO CON UNO STIMATORE NEURALE ITERATIVO DELL’ERRORE, CARATTERIZZATO DA PESI BINARIZZATI E ATTIVAZIONI A PUNTO FISSO. LA SOLUZIONE CONSENTE DI RIPRISTINARE L’ACCURATEZZA DEL SENSORE ENTRO ±0,5 HPA, RECUPERANDO FINO A 1,6 HPA, CON UNA POTENZA DINAMICA DELL’ORDINE DEI NANOWATT E UN’AREA DI 0,55 MM^2.
NEL LORO INSIEME, I RISULTATI DIMOSTRANO LA GENERALITÀ E L’EFFICACIA DEL FLUSSO DI PROGETTAZIONE PROPOSTO, MOSTRANDO COME, NONOSTANTE LA DIVERSITÀ DEI DOMINI APPLICATIVI E DEI REQUISITI PRESTAZIONALI, SIA POSSIBILE OTTENERE SOLUZIONI AD ALTA EFFICIENZA ENERGETICA E ACCURATEZZA COMPETITIVA. LA TESI DISTILLA INOLTRE PRINCIPI DI PROGETTAZIONE ISC GENERALIZZABILI, QUALI LA PROPAGAZIONE PRECOCE DEI VINCOLI, LA CONDIVISIONE AGGRESSIVA DELLE RISORSE CON CALCOLO SERIALIZZATO, LA QUANTIZZAZIONE A POCHI BIT E LA BINARIZZAZIONE SELETTIVA, NONCHÉ L’ADOZIONE DI MECCANISMI EVENT-TRIGGERED CON MODALITÀ DI DEEP SLEEP.
INFINE, DURANTE UN PERIODO DI RICERCA DI SEI MESI PRESSO LA JOHNS HOPKINS UNIVERSITY, È STATO CONDOTTO UNO STUDIO ESPLORATIVO SULL’IMPIEGO DI LARGE LANGUAGE MODELS (LLM) PER LA GENERAZIONE AUTOMATICA DI DESCRIZIONI HARDWARE, INCLUDENDO CODICE VERILOG SINTETIZZABILE, TESTBENCH E DOCUMENTAZIONE, APPLICATI ALLA PROGETTAZIONE DI UNA RETE NEURALE SPIKING RICORRENTE VALIDATA SU FPGA E IMPLEMENTATA TRAMITE UN FLUSSO OPEN-SOURCE IN TECNOLOGIA SKYWATER A 130 NM. TALE STUDIO FORNISCE UNA PROSPETTIVA COMPLEMENTARE SU COME IL WORKFLOW DI CO-PROGETTAZIONE PROPOSTO POSSA ESSERE ULTERIORMENTE ACCELERATO.THE EXPLOSIVE GROWTH OF IOT DEVICES DEMANDS ON-SENSOR INTELLIGENCE THAT IS ACCURATE AND RADICALLY ENERGY-EFFICIENT. THIS DISSERTATION INVESTIGATES IN-SENSOR COMPUTING (ISC) THROUGH A CONSTRAINTS-FIRST HARDWARE–SOFTWARE CO-DESIGN METHODOLOGY TO REALIZE TINY, ULTRA-LOW-POWER NEURAL ACCELERATORS TIGHTLY COUPLED TO MEMS SENSORS.
THE APPROACH SHAPES NETWORK TOPOLOGY, QUANTIZATION, AND FIXED-POINT ARITHMETIC FROM THE OUTSET TO MEET STRINGENT LIMITS IN AREA, POWER, AND MEMORY, AND VALIDATES DESIGNS THROUGH FPGA PROTOTYPING AND CMOS SYNTHESIS.
THREE APPLICATION-DRIVEN CASE STUDIES, DEVELOPED IN COLLABORATION WITH STMICROELECTRONICS, SUBSTANTIATE THE METHODOLOGY.
THE FIRST ADDRESSES AUDIO PROCESSING FOR KEYWORD SPOTTING, WHERE A LEARNED 1D-CNN REPLACES THE CONVENTIONAL CIC+FIR PDM-TO-PCM CHAIN, FUSING FILTERING AND DECIMATION AND DELIVERING 8-BIT/16 KHZ PCM WITH 48 DB SNR WHILE PRESERVING DOWNSTREAM ACCURACY OF 89%. THE SYNTHESIZED CORE IN 130 NM CMOS ACHIEVES 128.7 ΜW/MHZ WITHIN LESS THAN 1 MM^2.
THE SECOND FOCUSES ON VIBRATION-BASED PREDICTIVE MAINTENANCE, EMPLOYING A HYBRID, EVENT-DRIVEN PIPELINE THAT COMBINES AN ALWAYS-ON, PARTIALLY BINARIZED IN-SENSOR AUTOENCODER FOR ANOMALY DETECTION (AUC = 0.99; 99.61% ACCURACY) WITH AN ON-DEMAND MCU CLASSIFIER (UP TO 94.83%).
THE IN-SENSOR ACCELERATOR SUSTAINS SENSOR OUTPUT DATA RATES UP TO 365 KHZ AND EXHIBITS 333 ΜW/MHZ DYNAMIC POWER ON FPGA, WHILE STANDARD-CELL SYNTHESIS IN 65 NM REPORTS 0.49 MM^2 AND 138.6 ΜW/MHZ DYNAMIC POWER.
THE THIRD CASE CONCERNS THERMAL-STRESS COMPENSATION FOR MEMS PRESSURE SENSORS: THE PROPOSED AI-BASED RECONFIGURABLE SENSOR COMPENSATION UNIT (AI-RESCU) COUPLES A RECONFIGURABLE TRIGGER WITH AN ITERATIVE NEURAL ERROR ESTIMATOR WITH BINARIZED WEIGHTS AND FIXED-POINT ACTIVATIONS TO RESTORE ACCURACY WITHIN ±0.5 HPA, RECOVERING UP TO 1.6 HPA, WITH 4.46 NW DYNAMIC POWER IN 0.55 MM^2.
TAKEN TOGETHER, THESE DIVERSE STUDIES CONFIRM THE GENERAL APPLICABILITY OF THE PROPOSED DESIGN FLOW: DESPITE THEIR DIFFERENT SENSING DOMAINS AND PERFORMANCE TARGETS, EACH ACHIEVES STATE-OF-THE-ART ACCURACY AND EFFICIENCY.
THE DISSERTATION DISTILLS GENERALIZABLE ISC DESIGN PRINCIPLES—EARLY CONSTRAINT PROPAGATION, AGGRESSIVE RESOURCE SHARING WITH SERIALIZED COMPUTE, SELECTIVE BINARIZATION AND LOW-BIT QUANTIZATION, AND EVENT-TRIGGERED OPERATION WITH DEEP SLEEP—SHOWING THAT COMPETITIVE MACHINE-LEARNING ACCURACY AND REAL-TIME THROUGHPUT CAN BE ACHIEVED AT MILLIWATT-TO-NANOWATT POWER AND SUB-MM^2 AREA, ENABLING PRACTICAL IN-SENSOR AI. FINALLY, DURING A 6-MONTH RESEARCH PERIOD AT JOHNS HOPKINS UNIVERSITY, AN EXPLORATORY STUDY INVESTIGATED LARGE LANGUAGE MODELS (LLMS)-ASSISTED
HARDWARE-DESCRIPTION GENERATION, INCLUDING SYNTHESIZABLE VERILOG, TESTBENCHES, AND DOCUMENTATION FOR A RECURRENT SPIKING NEURAL NETWORK VALIDATED ON FPGA AND IMPLEMENTED WITH AN OPEN-SOURCE SKYWATER 130 NM FLOW, AS A COMPLEMENTARY PERSPECTIVE ON HOW THE PROPOSED CO-DESIGN WORKFLOW COULD BE ACCELERATED
VALIDATION-DRIVEN LLM ARCHITECTURES FOR CODE GENERATION, TRANSLATION, AND AUTOMATED GRADING
No full textNEGLI ULTIMI ANNI, I LARGE LANGUAGE MODELS (LLM) HANNO DIMOSTRATO NOTEVOLI CAPACITÀ NELL’ELABORAZIONE DEL LINGUAGGIO NATURALE E DEI LINGUAGGI DI PROGRAMMAZIONE. ADDESTRATI SU GRANDI QUANTITÀ DI DATI E CODICE, QUESTI MODELLI SONO IN GRADO DI COMPRENDERE E GENERARE TESTO CON ELEVATA ACCURATEZZA. TUTTAVIA, NONOSTANTE TALI CAPACITÀ, L’OUTPUT PRODOTTO DAI LLM NON È SEMPRE AFFIDABILE, IN PARTICOLARE NEL CONTESTO DEI LINGUAGGI DI PROGRAMMAZIONE E DEI LINGUAGGI FORMALI, DOVE LA PRECISIONE È UN REQUISITO FONDAMENTALE.
PER AFFRONTARE TALI LIMITAZIONI, LA LETTERATURA RECENTE PROPONE L’INTEGRAZIONE DI PROCESSI DI VALIDAZIONE ALL’INTERNO DELLE PIPELINE BASATE SU LLM, AL FINE DI VERIFICARE IL RISPETTO DI REQUISITI SINTATTICI E SEMANTICI SPECIFICI. IN QUESTO CONTESTO SI COLLOCANO I WORKFLOW GUIDATI DALLA VALIDAZIONE, CHE SFRUTTANO MECCANISMI DI FEEDBACK ITERATIVI PER MIGLIORARE PROGRESSIVAMENTE LA QUALITÀ DELL’OUTPUT GENERATO.
QUESTA TESI ANALIZZA FRAMEWORK E METODOLOGIE BASATI SULLA VALIDAZIONE CHE IMPIEGANO I LLM PER LA GENERAZIONE E LA TRADUZIONE DI LINGUAGGI DI PROGRAMMAZIONE, NONCHÉ LA LORO APPLICAZIONE IN AMBITO EDUCATIVO PER LA VALUTAZIONE AUTOMATIZZATA DEL CODICE. I WORKFLOW PROPOSTI INTEGRANO PROCEDURE DI VALIDAZIONE CHE GUIDANO LE SUCCESSIVE ITERAZIONI DI GENERAZIONE E TRADUZIONE, INCREMENTANDO L’ACCURATEZZA E L’AFFIDABILITÀ DEI RISULTATI.
NELLA PRIMA PARTE DELLA TESI VIENE STUDIATO L’UTILIZZO DI WORKFLOW GUIDATI DALLA VALIDAZIONE PER LA GENERAZIONE E LA TRADUZIONE DI DATI SEMI-STRUTTURATI E LINGUAGGI DI PROGRAMMAZIONE. IN PARTICOLARE, VIENE PRESENTATA UNA PIPELINE PER LA GENERAZIONE DI RISORSE JSON-FHIR A PARTIRE DA NARRAZIONI CLINICHE, BASATA SU UN CICLO DI FEEDBACK CHE UTILIZZA IL VALIDATORE FHIR UFFICIALE. VIENE INOLTRE ANALIZZATO UN WORKFLOW PER LA TRADUZIONE DI CODICE TRA PARADIGMI PROCEDURALI E FUNZIONALI IN JAVASCRIPT, CHE INTEGRA ANALISI STATICA E FEEDBACK DELL’INTERPRETE PER MIGLIORARE LA PRESERVAZIONE SEMANTICA. INFINE, VIENE CONSIDERATO UN WORKFLOW PER LA SINTESI DI PROGRAMMI IN FLUID, UN LINGUAGGIO ORIENTATO ALLA VISUALIZZAZIONE DEI DATI, MOSTRANDO COME STRATEGIE DI PROMPTING CONSAPEVOLI DELLA VERIFICA POSSANO MIGLIORARE LA CORRETTEZZA DEI PROGRAMMI GENERATI.
LA SECONDA PARTE DELLA TESI ESPLORA L’IMPIEGO DEI LLM IN CONTESTI EDUCATIVI PER LA VALUTAZIONE AUTOMATICA DEL CODICE. VIENE PROPOSTO UN FRAMEWORK PER LA VALUTAZIONE DELLE CONSEGNE DEGLI STUDENTI IN ESERCITAZIONI DI PROGRAMMAZIONE, SUCCESSIVAMENTE ESTESO INCLUDENDO INFORMAZIONI AGGIUNTIVE QUALI OUTPUT ATTESO, RISULTATI DEI TEST, FEEDBACK DEL COMPILATORE E ANALISI STATICA. I RISULTATI SPERIMENTALI MOSTRANO CHE I VOTI GENERATI DAI LLM DIFFERISCONO DA QUELLI UMANI DI CIRCA 1,15 PUNTI SU 10 E CHE L’UTILIZZO DI INFORMAZIONI AGGIUNTIVE MIGLIORA LE PRESTAZIONI DEI MODELLI.
INFINE, VIENE DISCUSSO L’IMPATTO DEI LLM IN AMBITO EDUCATIVO, CON PARTICOLARE ATTENZIONE AL FENOMENO DELL’OVERTRUST DA PARTE DEGLI STUDENTI, E VIENE ANALIZZATO L’IMPATTO AMBIENTALE DEI MODELLI PER IL CODICE, EVIDENZIANDO LE DIFFICOLTÀ DI STIMA DELL’IMPRONTA DI CARBONIO A PARTIRE DALLE INFORMAZIONI ATTUALMENTE DISPONIBILI IN LETTERATURA.IN RECENT YEARS, LARGE LANGUAGE MODELS (LLMS) HAVE EMERGED AS POWERFUL TOOLS FOR NATURAL AND PROGRAMMING LANGUAGE PRO- CESSING. LLMS ARE TRAINED ON VAST AMOUNTS OF TEXT AND CODE DATA, ENABLING THEM TO UNDERSTAND AND GENERATE HUMAN-LIKE TEXT WITH HIGH ACCURACY. HOWEVER, DESPITE THEIR CAPABILITIES, THE OUTPUTS GENERATED BY LLMS ARE NOT ALWAYS RELIABLE OR ACCURATE, ESPECIALLY IN THE CONTEXT OF PROGRAMMING AND FORMAL LANGUAGES, WHERE PRECISION IS CRUCIAL. TO ADDRESS THIS CHALLENGE, SOME APPROACHES HAVE PROPOSED THE INTEGRATION OF VALIDATION STEPS INTO LLM-BASED PIPELINES.THE VALIDATION OF LLM OUTPUTS INVOLVES PROCEDURES TO ENSURE COMPLIANCE WITH TASK-SPECIFIC REQUIREMENTS. IN MADAAN ET AL., FOR INSTANCE, THE LLM-BASED SYSTEM VERIFIES THE GENERATED OUTPUT THROUGH A NEW LLM CALL AND, IN CASE OF ERRORS, PROVIDES FEEDBACK TO THE MODEL TO REFINE ITS OUTPUT ITERATIVELY.OTHER APPROACHES INCORPORATE DIFFERENT VALIDATION MECHANISMS, BASED ON STATIC ANALYSIS, COMPILER DIAGNOSTICS, OR UNIT TESTS.THIS THESIS EXPLORES VARIOUS VALIDATION DRIVEN FRAMEWORKS AND METHODOLOGIES THAT UTILIZE LLMS FOR PROGRAMMING LANGUAGE GENERATION AND TRANSLATION, AS WELL AS THEIR DEPLOYMENT IN EDUCATIONAL CONTEXTS FOR AUTOMATED CODE EVALUATION. THESE VALIDATION-CENTRIC WORKFLOWS FREQUENTLY INTEGRATE ITERATIVE FEEDBACK MECHANISMS, WHERE THE OUTCOMES OF THE VALIDATION PROCESS GUIDE SUBSEQUENT ITERATIONS OF GENERATION OR TRANSLATION, THEREBY IMPROVING THE ACCURACY AND DEPENDABILITY OF LLM OUTPUTS. IN THE FIRST PART, WE PRESENT THE USE OF VALIDATION DRIVEN WORKFLOWS FOR THE GENERATION AND TRANSLATION OF SEMI- STRUCTURED DATA AND PROGRAMMING LANGUAGES. WE PRESENT A VALI- DATION DRIVEN PIPELINE FOR THE GENERATION OF JSON-FHIR RESOURCES FROM CLINICAL NARRATIVES, WHICH INCORPORATES A VALIDATION-FEEDBACK LOOP, BASED ON THE OFFICIAL FHIR VALIDATOR, TO ITERATIVELY REFINE THE GENERATED RESOURCES.THEN, WE PROPOSE A VALIDATION-AWARE WORKFLOW FOR TRANSLATING CODE BETWEEN PROCEDURAL AND FUNCTIONAL PARADIGMS IN JAVASCRIPT, INTEGRATING STATIC ANALYSIS AND FEEDBACK WITHIN THE PROMPTING STRATEGY TO ENHANCE SEMANTIC PRESERVATION DURING TRANSLATION.WE ALSO ANALYSE A VALIDATION DRIVEN WORKFLOW FOR SYNTHESIZING FLUID PROGRAMS, DEMONSTRATING THAT VERIFICATION-AWARE PROMPTING STRATEGIES CAN ENHANCE THE CORRECTNESS OF GENERATED PROGRAMS.FLUID IS A PROGRAMMING LANGUAGE DESIGNED FOR DATA VISUALISATION TASKS, PARTICULARLY FOCUSING ON TRANSPARENCY.IN THE SECOND PART OF THIS DISSERTATION, WE EXPLORE THE APPLICATION OF LLMS IN EDUCATIONAL CONTEXTS, SPECIFICALLY FOR AUTOMATED CODE GRADING AND ASSESSMENT.WE PROPOSE A FRAMEWORK THAT LEVERAGES LLMS TO AUTOMATICALLY EVALUATE STUDENT SUBMISSIONS IN PROGRAMMING ASSIGNMENTS, EMPIRICALLY ANALYSING DIFFERENT LLMS.WE THEN EXTEND THE WORKFLOW AND THE INFORMATION USED IN THE GRADING PHASE, PROVIDING NOT ONLY THE CODE BUT ALSO THE EXPECTED OUTPUT, THE TEST RESULTS, THE COMPILER FEEDBACK, AND STATIC ANALYSIS RESULTS.RESULTS SHOW THAT LLM-GENERATED GRADES DIFFER FROM HUMAN GRADES BY ABOUT 1.15 POINTS OUT OF 10, AND THAT THE INCLUSION OF ADDITIONAL INFORMATION DURING THE GRADING PHASE IMPROVES THE PERFORMANCE OF LLMS IN CODE ASSESSMENT TASKS.FINALLY, WE DISCUSS THE IMPACT OF LLMS IN EDUCATIONAL CONTEXTS, HIGHLIGHTING THE OVERTHRUST PHENOMENON AMONG STUDENTS, WHO TEND TO OVER- RELY ON LLM OUTPUTS INSTEAD OF USING THEM AS SUP- PORT TOOLS. MOREOVER, WE ALSO ANALYSE THE ENVIRONMENTAL IMPACT OF CODE MODELS, ASSESSING THE CARBON FOOTPRINT REPORTED IN THE LITERATURE.THE RESULTS SHOW THAT SEVERAL STUDIES DO NOT ATTEMPT TO ESTIMATE THEIR IMPACT AND GENERALLY REPORT ONLY TRAINING INFORMATION; WITH THIS INFORMATION IT IS DIFFICULT TO PRECISELY ESTIMATE THE IMPACT, ALTHOUGH WE PROVIDE APPROXIMATE ESTIMATES.FUTURE WORK INCLUDES EXTENDING THE ANALYSIS OF LLMS TO ADDITIONAL PROGRAMMING LANGUAGES, FORMAL LANGUAGES, AND PARADIGMS, INVESTIGATING NON-LLM AI MODELS FOR LANGUAGE-SPECIFIC TASKS, AND DEVELOPING RIGOROUS, EXPLAINABLE TECHNIQUES FOR AUTOMATED CODE GRADING AND ASSESSMENT
Analysing labour market indicators in entrepreneurship: an automated R framework for integrating bibliometric and narrative analysis
No full textThis study offers a comprehensive analysis of labour market indicators (LMIs) within the context of entrepreneurship, employing a hybrid methodology that combines systematic literature review (SLR), bibliometric techniques, and narrative content analysis. Drawing on a corpus of 242 peer-reviewed articles retrieved from Scopus and Web of Science. The research maps the rising of attention to the existing link between labour market dynamics and big data analytics. Utilizing the Biblioshiny interface in R, the study constructs co-citation and keyword co-occurrence networks, identifying thematic clusters and research trends from 2010 to 2025. Results reveal an expanding but fragmented field, with limited continuity among authors and underexplored dimensions in LMI evaluation. The analysis highlights three macro-themes: (i) technological innovation and its implications for labour markets, (ii) the evolving skill demands in terms of knowledge economy, and (iii) methodological approaches to labour policy evaluation. In particular, the study emphasizes the role of big data in mitigating informational asymmetries in employment matching and policy design. Additionally, it explores theoretical models related to skills mismatch, public policy effectiveness, and the impact of artificial intelligence on employment structures. The findings underscore the need for integrative frameworks that merge empirical rigor with conceptual depth, especially considering accelerating technological change and its socio-economic implications. This paper contributes to the literature by proposing a replicable methodological framework for mapping complex research fields, offering insights for scholars, policymakers, and practitioners concerned with labour market transformations in the digital age
RUTHENIUM-CATALYZED OLEFIN METATHESIS: APPLICATIONS IN GREEN AND SUSTAINABLE CHEMISTRY
No full textTHIS DOCTORAL THESIS FOCUSES ON THE PRIMARY OBJECTIVE OF DEVELOPING NEW HOVEYDA-GRUBBS TYPE RUTHENIUM CATALYSTS FOR OLEFIN METATHESIS FEATURING UNSYMMETRICAL N-BENZYL, N’-ARYL NHC FRAMEWORKS WITH SUBSTITUTED BACKBONES IN BOTH SYN AND ANTI CONFIGURATIONS. FOLLOWING THEIR SYNTHESIS, THESE NOVEL CATALYSTS WILL BE EVALUATED ACROSS A BROAD RANGE OF CATALYTIC APPLICATIONS INVOLVING STANDARD AND RENEWABLE SUBSTRATES, INCLUDING RING-CLOSING METATHESIS, CROSS METATHESIS, SELF-METATHESIS, ETHENOLYSIS AND RING-OPENING METATHESIS POLYMERIZATION. THE RESULTING PERFORMANCES WILL BE COMPARED WITH THOSE OF THE COMMERCIALLY AVAILABLE SYMMETRICAL HOVEYDA-GRUBBS SECOND GENERATION CATALYST (HGII) AND OTHER RELATED NHC CATALYSTS PREVIOUSLY REPORTED IN THE LITERATURE. OVERALL, THIS STUDY AIMS TO ELUCIDATE HOW SPECIFIC STRUCTURAL MODIFICATIONS TO NHC LIGANDS AFFECT CATALYST EFFICIENCY ACROSS DIVERSE METATHESIS TRANSFORMATIONS. AN ADDITIONAL AIM IS TO OPTIMIZE THE REACTION CONDITIONS FOR CROSS METATHESIS TO PROMOTE THE VALORIZATION OF WASTE-DERIVED Α-ALKENES OBTAINED FROM PLASTIC WASTE, PARTICULARLY POLYPROPYLENE AND POLYETHYLENE. THIS OPTIMIZATION PROCESS WILL INVOLVE SYSTEMATIC ADJUSTMENT OF REACTION PARAMETERS SUCH AS TEMPERATURE, CONCENTRATION, REACTION TIME, SOLVENT CHOICE, AND CATALYST ADDITION PROTOCOLS. FURTHERMORE, A COMPREHENSIVE INVESTIGATION OF THE SCOPE AND LIMITATIONS OF THE CROSS METATHESIS REACTION WILL BE CARRIED OUT USING A VARIETY OF CROSS-PARTNERS
Technologies for Supporting Academic Development
No full textThe entry analyses Academic Development as a strategic approach to enhancing university teaching in the digital era, focusing on educational technologies, theoretical frameworks such as TPACK, and innovative practices including blended learning and flipped classrooms. It also addresses key challenges and outlines evidence-based perspectives for strengthening teaching quality and institutional support
